diff --git a/internal/ubjdecode/decode.go b/internal/ubjdecode/decode.go
new file mode 100644
index 0000000..f9eb870
--- /dev/null
+++ b/internal/ubjdecode/decode.go
@@ -0,0 +1,536 @@
+// Package ubjdecode implements a standalone Universal Binary JSON (UBJSON) decoder.
+// Zero external dependencies — uses only encoding/binary and io from stdlib.
+//
+// UBJSON spec: https://ubjson.org/
+// XGBoost uses UBJSON for its .ubj model format.
+package ubjdecode
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"math"
+)
+
+// DecodeValue reads one UBJSON value from r and returns it as a native Go value:
+//
+//	map[string]interface{} for objects
+//	[]interface{} for untyped/mixed arrays
+//	[]int32, []int64, []float32, []float64 for optimized typed arrays
+//	string, bool, int64, float64, nil for scalars
+func DecodeValue(r io.Reader) (interface{}, error) {
+	marker, err := readByte(r)
+	if err != nil {
+		return nil, err
+	}
+	return decodeMarker(r, marker)
+}
+
+func decodeMarker(r io.Reader, marker byte) (interface{}, error) {
+	for marker == 'N' { // noop — skip and read next
+		var err error
+		marker, err = readByte(r)
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	switch marker {
+	case 'Z': // null
+		return nil, nil
+	case 'T': // true
+		return true, nil
+	case 'F': // false
+		return false, nil
+	case 'i': // int8
+		b, err := readByte(r)
+		if err != nil {
+			return nil, err
+		}
+		return int64(int8(b)), nil
+	case 'U': // uint8
+		b, err := readByte(r)
+		if err != nil {
+			return nil, err
+		}
+		return int64(b), nil
+	case 'I': // int16 big-endian
+		var v int16
+		if err := binary.Read(r, binary.BigEndian, &v); err != nil {
+			return nil, err
+		}
+		return int64(v), nil
+	case 'l': // int32 big-endian
+		var v int32
+		if err := binary.Read(r, binary.BigEndian, &v); err != nil {
+			return nil, err
+		}
+		return int64(v), nil
+	case 'L': // int64 big-endian
+		var v int64
+		if err := binary.Read(r, binary.BigEndian, &v); err != nil {
+			return nil, err
+		}
+		return v, nil
+	case 'd': // float32 big-endian
+		var bits uint32
+		if err := binary.Read(r, binary.BigEndian, &bits); err != nil {
+			return nil, err
+		}
+		return float64(math.Float32frombits(bits)), nil
+	case 'D': // float64 big-endian
+		var bits uint64
+		if err := binary.Read(r, binary.BigEndian, &bits); err != nil {
+			return nil, err
+		}
+		return math.Float64frombits(bits), nil
+	case 'C': // char — single byte as string
+		b, err := readByte(r)
+		if err != nil {
+			return nil, err
+		}
+		return string([]byte{b}), nil
+	case 'S': // string — length marker + bytes
+		n, err := readLength(r)
+		if err != nil {
+			return nil, fmt.Errorf("string length: %w", err)
+		}
+		buf := make([]byte, n)
+		if _, err := io.ReadFull(r, buf); err != nil {
+			return nil, err
+		}
+		return string(buf), nil
+	case 'H': // high-precision number — decode as string
+		n, err := readLength(r)
+		if err != nil {
+			return nil, fmt.Errorf("high-precision length: %w", err)
+		}
+		buf := make([]byte, n)
+		if _, err := io.ReadFull(r, buf); err != nil {
+			return nil, err
+		}
+		return string(buf), nil
+	case '[': // array
+		return decodeArray(r)
+	case '{': // object
+		return decodeObject(r)
+	default:
+		return nil, fmt.Errorf("ubjdecode: unknown marker byte 0x%02X ('%c')", marker, marker)
+	}
+}
+
+// readByte reads a single byte from r.
+func readByte(r io.Reader) (byte, error) {
+	var buf [1]byte
+	_, err := io.ReadFull(r, buf[:])
+	return buf[0], err
+}
+
+// readLength reads an integer length value (used for string lengths, array/object counts).
+// The next byte is a type marker ('i','U','I','l','L') indicating the integer width.
+func readLength(r io.Reader) (int64, error) {
+	marker, err := readByte(r)
+	if err != nil {
+		return 0, err
+	}
+	switch marker {
+	case 'i':
+		b, err := readByte(r)
+		return int64(int8(b)), err
+	case 'U':
+		b, err := readByte(r)
+		return int64(b), err
+	case 'I':
+		var v int16
+		err := binary.Read(r, binary.BigEndian, &v)
+		return int64(v), err
+	case 'l':
+		var v int32
+		err := binary.Read(r, binary.BigEndian, &v)
+		return int64(v), err
+	case 'L':
+		var v int64
+		err := binary.Read(r, binary.BigEndian, &v)
+		return v, err
+	default:
+		return 0, fmt.Errorf("ubjdecode: unexpected length marker 0x%02X ('%c')", marker, marker)
+	}
+}
+
+// decodeArray decodes a UBJSON array. The '[' marker has already been consumed.
+// Handles all four container forms:
+//  1. [$][type][#][len] — typed + counted → returns typed Go slice
+//  2. [#][len]          — counted only    → returns []interface{}
+//  3. [$][type] + ']'   — typed, terminated
+//  4. ']' terminator    — untyped, terminated → returns []interface{}
+func decodeArray(r io.Reader) (interface{}, error) {
+	next, err := readByte(r)
+	if err != nil {
+		return nil, err
+	}
+
+	// Typed array: [$][type][#][count]
+	if next == '$' {
+		typeMarker, err := readByte(r)
+		if err != nil {
+			return nil, err
+		}
+		// Peek for '#'
+		after, err := readByte(r)
+		if err != nil {
+			return nil, err
+		}
+		if after == '#' {
+			// typed + counted: fast path
+			count, err := readLength(r)
+			if err != nil {
+				return nil, fmt.Errorf("typed array count: %w", err)
+			}
+			return readTypedArray(r, typeMarker, int(count))
+		}
+		// typed + unterminated: read until ']'
+		// `after` is the first element's marker? No — in typed arrays the element
+		// marker is NOT repeated; `after` here should be ']' or data bytes.
+		// Actually when typed without count, each element has no marker prefix —
+		// the type is fixed. So `after` is the first data byte of the first element.
+		// We must decode accordingly.
+		return readTypedArrayUntilEnd(r, typeMarker, after)
+	}
+
+	// Counted only: [#][len]
+	if next == '#' {
+		count, err := readLength(r)
+		if err != nil {
+			return nil, fmt.Errorf("counted array length: %w", err)
+		}
+		result := make([]interface{}, 0, int(count))
+		for i := int64(0); i < count; i++ {
+			v, err := DecodeValue(r)
+			if err != nil {
+				return nil, fmt.Errorf("array element %d: %w", i, err)
+			}
+			result = append(result, v)
+		}
+		return result, nil
+	}
+
+	// Unterminated untyped: read until ']'
+	if next == ']' {
+		return []interface{}{}, nil
+	}
+	// next is the marker of the first element
+	result := []interface{}{}
+	v, err := decodeMarker(r, next)
+	if err != nil {
+		return nil, fmt.Errorf("array element 0: %w", err)
+	}
+	result = append(result, v)
+	for {
+		marker, err := readByte(r)
+		if err != nil {
+			return nil, err
+		}
+		if marker == ']' {
+			break
+		}
+		v, err := decodeMarker(r, marker)
+		if err != nil {
+			return nil, fmt.Errorf("array element: %w", err)
+		}
+		result = append(result, v)
+	}
+	return result, nil
+}
+
+// readTypedArray reads `count` elements of a fixed type without per-element markers.
+// Returns typed Go slices for performance-critical XGBoost arrays.
+func readTypedArray(r io.Reader, typeMarker byte, count int) (interface{}, error) {
+	switch typeMarker {
+	case 'i': // int8 → []int32
+		buf := make([]int8, count)
+		for i := range buf {
+			b, err := readByte(r)
+			if err != nil {
+				return nil, err
+			}
+			buf[i] = int8(b)
+		}
+		out := make([]int32, count)
+		for i, v := range buf {
+			out[i] = int32(v)
+		}
+		return out, nil
+	case 'U': // uint8 → []int32
+		buf := make([]byte, count)
+		if _, err := io.ReadFull(r, buf); err != nil {
+			return nil, err
+		}
+		out := make([]int32, count)
+		for i, v := range buf {
+			out[i] = int32(v)
+		}
+		return out, nil
+	case 'I': // int16 → []int32
+		out := make([]int32, count)
+		for i := range out {
+			var v int16
+			if err := binary.Read(r, binary.BigEndian, &v); err != nil {
+				return nil, err
+			}
+			out[i] = int32(v)
+		}
+		return out, nil
+	case 'l': // int32 → []int32
+		out := make([]int32, count)
+		for i := range out {
+			if err := binary.Read(r, binary.BigEndian, &out[i]); err != nil {
+				return nil, err
+			}
+		}
+		return out, nil
+	case 'L': // int64 → []int64
+		out := make([]int64, count)
+		for i := range out {
+			if err := binary.Read(r, binary.BigEndian, &out[i]); err != nil {
+				return nil, err
+			}
+		}
+		return out, nil
+	case 'd': // float32 → []float32
+		out := make([]float32, count)
+		for i := range out {
+			var bits uint32
+			if err := binary.Read(r, binary.BigEndian, &bits); err != nil {
+				return nil, err
+			}
+			out[i] = math.Float32frombits(bits)
+		}
+		return out, nil
+	case 'D': // float64 → []float64
+		out := make([]float64, count)
+		for i := range out {
+			var bits uint64
+			if err := binary.Read(r, binary.BigEndian, &bits); err != nil {
+				return nil, err
+			}
+			out[i] = math.Float64frombits(bits)
+		}
+		return out, nil
+	default:
+		// Fall back to []interface{} for other typed arrays
+		result := make([]interface{}, 0, count)
+		for i := 0; i < count; i++ {
+			v, err := decodeMarker(r, typeMarker)
+			if err != nil {
+				return nil, fmt.Errorf("typed array element %d: %w", i, err)
+			}
+			result = append(result, v)
+		}
+		return result, nil
+	}
+}
+
+// readTypedArrayUntilEnd reads a typed array without a count, terminated by ']'.
+// firstDataByte is the first raw data byte already consumed (not a marker).
+func readTypedArrayUntilEnd(r io.Reader, typeMarker byte, firstDataByte byte) (interface{}, error) {
+	// For typed arrays without count, we need to reconstruct the first element
+	// and then keep reading until we see ']'.
+	// Strategy: build []interface{} since we don't know the count.
+	result := []interface{}{}
+
+	// Decode first element from firstDataByte
+	first, err := decodeSingleFromByte(typeMarker, firstDataByte, r)
+	if err != nil {
+		return nil, err
+	}
+	result = append(result, first)
+
+	for {
+		// For typed arrays, next byte is raw data (no marker), or ']' to end
+		b, err := readByte(r)
+		if err != nil {
+			return nil, err
+		}
+		if b == ']' {
+			break
+		}
+		v, err := decodeSingleFromByte(typeMarker, b, r)
+		if err != nil {
+			return nil, err
+		}
+		result = append(result, v)
+	}
+	return result, nil
+}
+
+// decodeSingleFromByte decodes one element of a typed array given the first
+// byte already consumed and any remaining bytes read from r.
+func decodeSingleFromByte(typeMarker byte, firstByte byte, r io.Reader) (interface{}, error) {
+	switch typeMarker {
+	case 'i':
+		return int64(int8(firstByte)), nil
+	case 'U':
+		return int64(firstByte), nil
+	case 'I':
+		var buf [1]byte
+		if _, err := io.ReadFull(r, buf[:]); err != nil {
+			return nil, err
+		}
+		v := int16(firstByte)<<8 | int16(buf[0])
+		return int64(v), nil
+	case 'l':
+		var rest [3]byte
+		if _, err := io.ReadFull(r, rest[:]); err != nil {
+			return nil, err
+		}
+		v := int32(firstByte)<<24 | int32(rest[0])<<16 | int32(rest[1])<<8 | int32(rest[2])
+		return int64(v), nil
+	case 'L':
+		var rest [7]byte
+		if _, err := io.ReadFull(r, rest[:]); err != nil {
+			return nil, err
+		}
+		v := int64(firstByte)<<56 | int64(rest[0])<<48 | int64(rest[1])<<40 |
+			int64(rest[2])<<32 | int64(rest[3])<<24 | int64(rest[4])<<16 |
+			int64(rest[5])<<8 | int64(rest[6])
+		return v, nil
+	case 'd':
+		var rest [3]byte
+		if _, err := io.ReadFull(r, rest[:]); err != nil {
+			return nil, err
+		}
+		bits := uint32(firstByte)<<24 | uint32(rest[0])<<16 | uint32(rest[1])<<8 | uint32(rest[2])
+		return float64(math.Float32frombits(bits)), nil
+	case 'D':
+		var rest [7]byte
+		if _, err := io.ReadFull(r, rest[:]); err != nil {
+			return nil, err
+		}
+		bits := uint64(firstByte)<<56 | uint64(rest[0])<<48 | uint64(rest[1])<<40 |
+			uint64(rest[2])<<32 | uint64(rest[3])<<24 | uint64(rest[4])<<16 |
+			uint64(rest[5])<<8 | uint64(rest[6])
+		return math.Float64frombits(bits), nil
+	default:
+		// treat firstByte as the marker
+		return decodeMarker(r, firstByte)
+	}
+}
+
+// decodeObject decodes a UBJSON object. The '{' marker has already been consumed.
+// Returns map[string]interface{}.
+func decodeObject(r io.Reader) (map[string]interface{}, error) {
+	result := make(map[string]interface{})
+
+	// Check for optimized form: [{][#][len] or [{][$][type][#][len]
+	next, err := readByte(r)
+	if err != nil {
+		return nil, err
+	}
+
+	if next == '#' {
+		// counted object
+		count, err := readLength(r)
+		if err != nil {
+			return nil, fmt.Errorf("object count: %w", err)
+		}
+		for i := int64(0); i < count; i++ {
+			k, v, err := readKeyValue(r)
+			if err != nil {
+				return nil, fmt.Errorf("object key/value %d: %w", i, err)
+			}
+			result[k] = v
+		}
+		return result, nil
+	}
+
+	if next == '}' {
+		return result, nil
+	}
+
+	// Unoptimized: next byte is the length marker of the first key
+	// In UBJSON objects, keys are NOT prefixed with 'S' — just length_marker + bytes
+	k, v, err := readKeyValueFromLenMarker(r, next)
+	if err != nil {
+		return nil, fmt.Errorf("object first key/value: %w", err)
+	}
+	result[k] = v
+
+	for {
+		marker, err := readByte(r)
+		if err != nil {
+			return nil, err
+		}
+		if marker == '}' {
+			break
+		}
+		k, v, err := readKeyValueFromLenMarker(r, marker)
+		if err != nil {
+			return nil, fmt.Errorf("object key/value: %w", err)
+		}
+		result[k] = v
+	}
+	return result, nil
+}
+
+// readKeyValue reads one key-value pair from an object.
+// The key length marker has NOT been consumed yet.
+func readKeyValue(r io.Reader) (string, interface{}, error) {
+	lenMarker, err := readByte(r)
+	if err != nil {
+		return "", nil, err
+	}
+	return readKeyValueFromLenMarker(r, lenMarker)
+}
+
+// readKeyValueFromLenMarker reads one key-value pair where lenMarker is the
+// already-consumed first byte of the key length.
+func readKeyValueFromLenMarker(r io.Reader, lenMarker byte) (string, interface{}, error) {
+	// lenMarker is the type marker for the key length
+	var keyLen int64
+	switch lenMarker {
+	case 'i':
+		b, err := readByte(r)
+		if err != nil {
+			return "", nil, err
+		}
+		keyLen = int64(int8(b))
+	case 'U':
+		b, err := readByte(r)
+		if err != nil {
+			return "", nil, err
+		}
+		keyLen = int64(b)
+	case 'I':
+		var v int16
+		if err := binary.Read(r, binary.BigEndian, &v); err != nil {
+			return "", nil, err
+		}
+		keyLen = int64(v)
+	case 'l':
+		var v int32
+		if err := binary.Read(r, binary.BigEndian, &v); err != nil {
+			return "", nil, err
+		}
+		keyLen = int64(v)
+	case 'L':
+		var v int64
+		if err := binary.Read(r, binary.BigEndian, &v); err != nil {
+			return "", nil, err
+		}
+		keyLen = v
+	default:
+		return "", nil, fmt.Errorf("ubjdecode: unexpected key length marker 0x%02X ('%c')", lenMarker, lenMarker)
+	}
+
+	keyBuf := make([]byte, keyLen)
+	if _, err := io.ReadFull(r, keyBuf); err != nil {
+		return "", nil, fmt.Errorf("reading key: %w", err)
+	}
+	key := string(keyBuf)
+
+	val, err := DecodeValue(r)
+	if err != nil {
+		return "", nil, fmt.Errorf("value for key %q: %w", key, err)
+	}
+	return key, val, nil
+}
diff --git a/internal/xgjson/detect.go b/internal/xgjson/detect.go
new file mode 100644
index 0000000..3250ded
--- /dev/null
+++ b/internal/xgjson/detect.go
@@ -0,0 +1,121 @@
+package xgjson
+
+import (
+	"bufio"
+	"encoding/binary"
+	"bytes"
+)
+
+// LooksLikeJSON peeks at the buffered reader and returns true if the content
+// appears to be an XGBoost JSON model file. It does not advance the reader.
+//
+// Heuristic: the first non-whitespace byte must be '{', and the string
+// "\"learner\"" (the mandatory top-level key in every XGBoost model) must
+// appear within the first 512 peeked bytes.
+func LooksLikeJSON(r *bufio.Reader) (bool, error) {
+	buf, err := r.Peek(512)
+	if len(buf) == 0 {
+		return false, err
+	}
+
+	// Find first non-whitespace byte; it must be '{'
+	firstNonWS := -1
+	for i, b := range buf {
+		if b != ' ' && b != '\t' && b != '\r' && b != '\n' {
+			firstNonWS = i
+			break
+		}
+	}
+	if firstNonWS < 0 || buf[firstNonWS] != '{' {
+		return false, nil
+	}
+
+	// Confirm the mandatory top-level key is present in the peeked window.
+	// Every XGBoost JSON model has "learner" as its first object key.
+	return bytes.Contains(buf, []byte(`"learner"`)), nil
+}
+
+// LooksLikeUBJ peeks at the buffered reader and returns true if the content
+// appears to be an XGBoost UBJ (Universal Binary JSON) model file. It does
+// not advance the reader.
+//
+// Heuristic: UBJ objects start with '{' (0x7B), and the first key in an
+// XGBoost UBJ model is always "learner". We verify the first byte is '{' and
+// then parse the UBJ key-length encoding + key bytes from the peek buffer,
+// checking that the first key is exactly "learner".
+//
+// UBJ key lengths are encoded as: <type-marker> <integer-bytes>, where the
+// marker is one of: 'i' (int8), 'U' (uint8), 'I' (int16 BE), 'l' (int32 BE),
+// 'L' (int64 BE). XGBoost currently uses 'L', so we handle all widths for
+// robustness.
+func LooksLikeUBJ(r *bufio.Reader) (bool, error) {
+	// 32 bytes is more than enough: 1 ('{') + 1 (marker) + 8 (int64) + 7 ("learner") = 17
+	buf, err := r.Peek(32)
+	if len(buf) < 2 {
+		return false, err
+	}
+
+	if buf[0] != '{' {
+		return false, nil
+	}
+
+	// Parse the key length from buf[1:]
+	keyLen, headerLen, ok := ubjReadLength(buf[1:])
+	if !ok {
+		return false, nil
+	}
+
+	// "learner" is 7 bytes
+	if keyLen != 7 {
+		return false, nil
+	}
+
+	start := 1 + headerLen
+	end := start + 7
+	if end > len(buf) {
+		return false, nil
+	}
+
+	return bytes.Equal(buf[start:end], []byte("learner")), nil
+}
+
+// ubjReadLength parses a UBJ integer-length from b (starting at b[0] which is
+// the type marker). Returns (value, bytesConsumed, ok).
+func ubjReadLength(b []byte) (value int, consumed int, ok bool) {
+	if len(b) < 1 {
+		return 0, 0, false
+	}
+	marker := b[0]
+	switch marker {
+	case 'i': // int8 — 1 data byte
+		if len(b) < 2 {
+			return 0, 0, false
+		}
+		return int(int8(b[1])), 2, true
+	case 'U': // uint8 — 1 data byte
+		if len(b) < 2 {
+			return 0, 0, false
+		}
+		return int(b[1]), 2, true
+	case 'I': // int16 BE — 2 data bytes
+		if len(b) < 3 {
+			return 0, 0, false
+		}
+		v := int16(binary.BigEndian.Uint16(b[1:3]))
+		return int(v), 3, true
+	case 'l': // int32 BE — 4 data bytes
+		if len(b) < 5 {
+			return 0, 0, false
+		}
+		v := int32(binary.BigEndian.Uint32(b[1:5]))
+		return int(v), 5, true
+	case 'L': // int64 BE — 8 data bytes
+		if len(b) < 9 {
+			return 0, 0, false
+		}
+		v := binary.BigEndian.Uint64(b[1:9])
+		return int(v), 9, true
+	default:
+		return 0, 0, false
+	}
+}
diff --git a/internal/xgjson/detect_test.go b/internal/xgjson/detect_test.go
new file mode 100644
index 0000000..d5ec3ab
--- /dev/null
+++ b/internal/xgjson/detect_test.go
@@ -0,0 +1,110 @@
+package xgjson_test
+
+import (
+	"bufio"
+	"bytes"
+	"os"
+	"testing"
+
+	"github.com/dmitryikh/leaves/internal/xgjson"
+)
+
+const testdataDir = "testdata/"
+
+func TestLooksLikeJSON(t *testing.T) {
+	t.Run("real JSON model", func(t *testing.T) {
+		r := openBuf(t, testdataDir+"test_binary_logistic.json")
+		if ok, err := xgjson.LooksLikeJSON(r); err != nil || !ok {
+			t.Errorf("LooksLikeJSON(json) = (%v, %v), want (true, nil)", ok, err)
+		}
+	})
+
+	t.Run("UBJ file is not JSON", func(t *testing.T) {
+		r := openBuf(t, testdataDir+"test_binary_logistic.ubj")
+		if ok, _ := xgjson.LooksLikeJSON(r); ok {
+			t.Error("LooksLikeJSON(ubj) = true, want false")
+		}
+	})
+
+	t.Run("empty reader", func(t *testing.T) {
+		r := bufio.NewReader(bytes.NewReader(nil))
+		if ok, _ := xgjson.LooksLikeJSON(r); ok {
+			t.Error("LooksLikeJSON(empty) = true, want false")
+		}
+	})
+
+	t.Run("random bytes", func(t *testing.T) {
+		r := bufio.NewReader(bytes.NewReader([]byte{0x00, 0x01, 0x02, 0x03, 0xFF}))
+		if ok, _ := xgjson.LooksLikeJSON(r); ok {
+			t.Error("LooksLikeJSON(random bytes) = true, want false")
+		}
+	})
+
+	t.Run("reader is not consumed", func(t *testing.T) {
+		f, err := os.Open(testdataDir + "test_binary_logistic.json")
+		if err != nil {
+			t.Fatal(err)
+		}
+		defer f.Close()
+		r := bufio.NewReader(f)
+		xgjson.LooksLikeJSON(r) //nolint:errcheck
+		// Peek must not consume: a second call should return the same result.
+		if ok, _ := xgjson.LooksLikeJSON(r); !ok {
+			t.Error("second LooksLikeJSON call returned false — reader was consumed")
+		}
+	})
+}
+
+func TestLooksLikeUBJ(t *testing.T) {
+	t.Run("real UBJ model", func(t *testing.T) {
+		r := openBuf(t, testdataDir+"test_binary_logistic.ubj")
+		if ok, err := xgjson.LooksLikeUBJ(r); err != nil || !ok {
+			t.Errorf("LooksLikeUBJ(ubj) = (%v, %v), want (true, nil)", ok, err)
+		}
+	})
+
+	t.Run("JSON file is not UBJ", func(t *testing.T) {
+		r := openBuf(t, testdataDir+"test_binary_logistic.json")
+		if ok, _ := xgjson.LooksLikeUBJ(r); ok {
+			t.Error("LooksLikeUBJ(json) = true, want false")
+		}
+	})
+
+	t.Run("empty reader", func(t *testing.T) {
+		r := bufio.NewReader(bytes.NewReader(nil))
+		if ok, _ := xgjson.LooksLikeUBJ(r); ok {
+			t.Error("LooksLikeUBJ(empty) = true, want false")
+		}
+	})
+
+	t.Run("random bytes", func(t *testing.T) {
+		r := bufio.NewReader(bytes.NewReader([]byte{0x00, 0x01, 0x02, 0x03, 0xFF}))
+		if ok, _ := xgjson.LooksLikeUBJ(r); ok {
+			t.Error("LooksLikeUBJ(random bytes) = true, want false")
+		}
+	})
+
+	t.Run("reader is not consumed", func(t *testing.T) {
+		f, err := os.Open(testdataDir + "test_binary_logistic.ubj")
+		if err != nil {
+			t.Fatal(err)
+		}
+		defer f.Close()
+		r := bufio.NewReader(f)
+		xgjson.LooksLikeUBJ(r) //nolint:errcheck
+		// Peek must not consume: a second call should return the same result.
+		if ok, _ := xgjson.LooksLikeUBJ(r); !ok {
+			t.Error("second LooksLikeUBJ call returned false — reader was consumed")
+		}
+	})
+}
+
+func openBuf(t *testing.T, path string) *bufio.Reader {
+	t.Helper()
+	f, err := os.Open(path)
+	if err != nil {
+		t.Fatalf("open %s: %v", path, err)
+	}
+	t.Cleanup(func() { f.Close() })
+	return bufio.NewReader(f)
+}
diff --git a/internal/xgjson/testdata/README.md b/internal/xgjson/testdata/README.md
new file mode 100644
index 0000000..105be25
--- /dev/null
+++ b/internal/xgjson/testdata/README.md
@@ -0,0 +1,22 @@
+# XGBoost Test Model Data
+
+This directory contains generated XGBoost model files used by the `xgjson` package tests.
+
+## Files
+
+- `train.py` — Python script that generates the test models
+- `test1.json` — XGBoost model in JSON format (XGBoost 2.x/3.x)
+- `test1.ubj` — XGBoost model in Universal Binary JSON format (XGBoost 2.x/3.x)
+- `test1_expected.json` — Input feature vectors with expected raw scores and probabilities
+
+## Regenerating
+
+Requires [uv](https://docs.astral.sh/uv/).
+
+```sh
+cd internal/xgjson/testdata
+go generate
+```
+
+This trains a small XGBoost binary classifier (10 estimators, max_depth=3) on synthetic data
+with 3 features, then writes the three output files above. Commit all three output files.
diff --git a/internal/xgjson/testdata/generate.go b/internal/xgjson/testdata/generate.go
new file mode 100644
index 0000000..73179f5
--- /dev/null
+++ b/internal/xgjson/testdata/generate.go
@@ -0,0 +1,3 @@
+//go:generate uv run train.py
+
+package testdata
diff --git a/internal/xgjson/testdata/test_binary_logistic.json b/internal/xgjson/testdata/test_binary_logistic.json
new file mode 100644
index 0000000..8d20caf
--- /dev/null
+++ b/internal/xgjson/testdata/test_binary_logistic.json
@@ -0,0 +1 @@
+{"learner":{"attributes":{},"feature_names":[],"feature_types":[],"gradient_booster":{"model":{"cats":{"enc":[],"feature_segments":[],"sorted_idx":[]},"gbtree_model_param":{"num_parallel_tree":"1","num_trees":"10"},"iteration_indptr":[0,1,2,3,4,5,6,7,8,9,10],"tree_info":[0,0,0,0,0,0,0,0,0,0],"trees":[{"base_weights":[3.4848927E-8,-1.5204959E0,1.6623036E0,1.1777768E0,-1.6315176E0,7.10179E-1,1.8392156E0,4.33977E-1,7.083949E-2,-5.601467E-1,-1.6758814E-1,3.6204177E-1,-1.5752013E-1,4.071052E-1,5.7988805E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":0,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[5.1048126E2,3.3520996E1,1.4945984E1,8.6061764E-1,2.4676788E1,1.0907537E1,1.0265808E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[2.733439E-1,-1.8622892E0,5.279944E-1,1.3041685E0,-2.5623593E-1,1.3718903E0,7.7473426E-1,4.33977E-1,7.083949E-2,-5.601467E-1,-1.6758814E-1,3.6204177E-1,-1.5752013E-1,4.071052E-1,5.7988805E-1],"split_indices":[2,0,2,1,2,1,2,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.9996877E2,1.0448368E2,9.5485085E1,3.7494144E0,1.0073427E2,1.574754E1,7.973755E1,2.4996095E0,1.2498047E0,8.19872E1,1.8747072E1,1.1248243E1,4.499297E0,1.549758E1,6.423997E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[3.3946685E-3,-1.1543779E0,1.2591227E0,9.5409256E-1,-1.2454361E0,5.2215135E-1,1.4025306E0,3.727428E-1,2.3597144E-2,-4.3348977E-1,-1.21060215E-1,4.5374906E-1,2.7029645E-2,3.0377525E-1,4.4462258E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":1,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.7547772E2,2.011644E1,8.874084E0,1.0756211E0,1.5365356E1,6.8806324E0,8.410034E-1,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[2.733439E-1,-1.8622892E0,5.279944E-1,-7.9844646E-2,-2.5623593E-1,-1.8622892E0,7.7473426E-1,3.727428E-1,2.3597144E-2,-4.3348977E-1,-1.21060215E-1,4.5374906E-1,2.7029645E-2,3.0377525E-1,4.4462258E-1],"split_indices":[2,0,2,2,2,0,2,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.874821E2,9.758766E1,8.989444E1,3.647652E0,9.394001E1,1.5399192E1,7.449525E1,2.428422E0,1.2192299E0,7.53628E1,1.8577206E1,3.9179015E0,1.1481291E1,1.4947725E1,5.954752E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[7.773815E-3,-9.943017E-1,9.9700433E-1,2.7761754E-1,-1.076667E0,5.286407E-1,3.7626263E-1,-3.394221E-1,3.191901E-1,2.0238668E-1,-3.8180283E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0],"id":2,"left_children":[1,3,5,-1,7,9,-1,-1,-1,-1,-1],"loss_changes":[1.6563177E2,1.4121422E1,9.465561E0,0E0,1.0584633E1,9.1505995E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5],"right_children":[2,4,6,-1,8,10,-1,-1,-1,-1,-1],"split_conditions":[2.0596176E-1,-1.8622892E0,7.2267E-1,2.7761754E-1,-4.3772988E-2,2.3320758E0,3.7626263E-1,-3.394221E-1,3.191901E-1,2.0238668E-1,-3.8180283E-1],"split_indices":[2,0,2,0,0,1,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.6508813E2,8.200907E1,8.307906E1,2.9528005E0,7.905627E1,3.0533514E1,5.254555E1,7.754477E1,1.5115029E0,2.8823126E1,1.7103883E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"11","size_leaf_vector":"1"}},{"base_weights":[9.2651285E-3,-1.0554368E0,6.792614E-1,2.3023954E-1,-1.1098394E0,1.0147718E-1,1.0298864E0,-5.1912617E-2,-3.520825E-1,1.7861223E-1,-3.123424E-1,1.9472034E-1,3.3861062E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":3,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[1.0307969E2,6.3869324E0,1.786119E1,0E0,3.0589828E0,2.0146532E1,1.3401375E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-2.5623593E-1,-2.0155988E0,5.279944E-1,2.3023954E-1,-2.6036727E-1,1.3718903E0,7.7473426E-1,-5.1912617E-2,-3.520825E-1,1.7861223E-1,-3.123424E-1,1.9472034E-1,3.3861062E-1],"split_indices":[2,0,2,0,1,1,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.4249661E2,5.4814583E1,8.768203E1,1.2086432E0,5.3605938E1,3.360115E1,5.4080883E1,3.6531324E0,4.9952805E1,2.379479E1,9.806361E0,1.2701178E1,4.1379704E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[6.5199593E-3,-9.5052385E-1,5.596426E-1,2.0603801E-1,-1.0059912E0,7.9649486E-2,9.219858E-1,-3.384241E-1,-1.2668663E-1,1.2908778E-1,-2.4213417E-1,1.5698004E-1,3.0862683E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":4,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[6.568722E1,4.8187065E0,1.3602236E1,0E0,2.6854439E0,1.1204861E1,1.3651428E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-2.5623593E-1,-2.0155988E0,5.594684E-1,2.0603801E-1,-5.7463634E-1,1.3718903E0,7.7473426E-1,-3.384241E-1,-1.2668663E-1,1.2908778E-1,-2.4213417E-1,1.5698004E-1,3.0862683E-1],"split_indices":[2,0,2,0,2,1,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2208248E2,4.445151E1,7.763097E1,1.0835884E0,4.3367924E1,3.391881E1,4.371216E1,3.5094765E1,8.273158E0,2.4676153E1,9.242659E0,1.0479339E1,3.323282E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[3.5492254E-3,-5.9674454E-1,6.862838E-1,-1.0244098E0,-1.5051265E-1,3.5410693E-1,9.457186E-1,-1.983951E-2,-3.1750667E-1,-1.2765785E-1,4.3486968E-1,3.8029623E-1,1.0213314E-2,2.9135826E-1,1.1577517E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":5,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[4.385949E1,1.0713873E1,3.979826E0,8.703499E-1,1.3396497E1,6.8439E0,1.7572403E-2,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[3.263706E-1,-5.7463634E-1,7.7473426E-1,-1.534747E0,-7.9241884E-1,-1.8355771E0,1.0084084E0,-1.983951E-2,-3.1750667E-1,-1.2765785E-1,4.3486968E-1,3.8029623E-1,1.0213314E-2,2.9135826E-1,1.1577517E-1],"split_indices":[2,2,2,0,0,0,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.0501652E2,5.5949192E1,4.9067326E1,2.7907331E1,2.804186E1,2.255417E1,2.6513155E1,1.0556241E0,2.6851707E1,2.455545E1,3.4864097E0,5.08358E0,1.747059E1,2.4656202E1,1.8569524E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[9.643211E-3,-9.6156096E-1,3.2266313E-1,-1.6817693E-2,-1.0025502E0,1.896752E-2,3.062792E-1,-3.163531E-2,-3.1356356E-1,3.1077072E-1,-8.825563E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0],"id":6,"left_children":[1,3,5,-1,7,9,-1,-1,-1,-1,-1],"loss_changes":[2.8589937E1,7.9467964E-1,1.5057805E1,0E0,7.268925E-1,1.6408224E1,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5],"right_children":[2,4,6,-1,8,10,-1,-1,-1,-1,-1],"split_conditions":[-5.7463634E-1,-1.534747E0,-7.9241884E-1,-1.6817693E-2,-2.3389478E0,-1.8355771E0,3.062792E-1,-3.163531E-2,-3.1356356E-1,3.1077072E-1,-8.825563E-2],"split_indices":[2,0,0,0,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[9.202333E1,2.1924986E1,7.009835E1,1.052031E0,2.0872955E1,4.9567326E1,2.0531021E1,1.1051567E0,1.97678E1,1.1126637E1,3.844069E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"11","size_leaf_vector":"1"}},{"base_weights":[1.0875948E-2,-9.042081E-1,2.6562464E-1,-1.4255263E-2,-9.5357305E-1,2.432597E-2,9.072643E-1,-2.707678E-2,-3.0165362E-1,1.6966063E-1,-1.2472395E-1,2.841285E-1,9.571609E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":7,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[1.9578981E1,7.282772E-1,1.0144684E1,0E0,6.824322E-1,1.1845299E1,1.1762619E-1,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-5.7463634E-1,-1.534747E0,-7.9241884E-1,-1.4255263E-2,-2.3389478E0,-1.451989E0,1.0084084E0,-2.707678E-2,-3.0165362E-1,1.6966063E-1,-1.2472395E-1,2.841285E-1,9.571609E-2],"split_indices":[2,0,0,0,2,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[8.195767E1,1.7292343E1,6.466533E1,1.0490721E0,1.6243273E1,4.7747166E1,1.6918161E1,1.0771346E0,1.5166138E1,2.1285168E1,2.6462E1,1.5273769E1,1.6443915E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[8.420152E-3,-8.866207E-1,2.0744449E-1,-2.3173474E-2,-2.8409058E-1,3.5733518E-1,-6.553415E-1,3.0889487E-1,2.7870974E-2,2.424394E-1,-3.562318E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0],"id":8,"left_children":[1,3,5,-1,-1,7,9,-1,-1,-1,-1],"loss_changes":[1.3613471E1,6.0597706E-1,8.315063E0,0E0,0E0,9.483938E0,8.436519E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,5,5,6,6],"right_children":[2,4,6,-1,-1,8,10,-1,-1,-1,-1],"split_conditions":[-6.423459E-1,-2.3389478E0,1.8273275E0,-2.3173474E-2,-2.8409058E-1,-1.4220484E0,-2.035454E0,3.0889487E-1,2.7870974E-2,2.424394E-1,-3.562318E-1],"split_indices":[2,2,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[7.438975E1,1.2903961E1,6.148579E1,1.053575E0,1.1850386E1,5.2884132E1,8.601657E0,1.4111651E1,3.8772484E1,2.1552174E0,6.4464393E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"11","size_leaf_vector":"1"}},{"base_weights":[3.1257947E-3,-2.370009E-1,7.368995E-1,-4.209105E-5,-8.483909E-1,8.1515753E-1,1.465431E-1,3.7057182E-1,-1.14696495E-1,2.5674397E-1,-3.5496542E-1,7.9826936E-2,2.756763E-1,-1.3072331E-1,1.8127826E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":9,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.2430462E1,7.701972E0,6.841221E-1,1.9037214E1,9.485901E0,6.1796856E-1,1.1308331E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[7.2267E-1,1.398677E0,1.8586534E0,-1.2937937E0,-2.0780978E0,6.0014886E-1,1.2276509E0,3.7057182E-1,-1.14696495E-1,2.5674397E-1,-3.5496542E-1,7.9826936E-2,2.756763E-1,-1.3072331E-1,1.8127826E-1],"split_indices":[2,1,1,0,0,1,2,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[6.855681E1,5.2163403E1,1.6393408E1,3.8313972E1,1.3849428E1,1.4138422E1,2.254987E0,8.527461E0,2.9786512E1,2.0186934E0,1.1830735E1,2.8137515E0,1.132467E1,1.0135728E0,1.2414142E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}}]},"name":"gbtree"},"learner_model_param":{"base_score":"[4.9375E-1]","boost_from_average":"1","num_class":"0","num_feature":"3","num_target":"1"},"objective":{"name":"binary:logistic","reg_loss_param":{"scale_pos_weight":"1"}}},"version":[3,2,0]}
\ No newline at end of file
diff --git a/internal/xgjson/testdata/test_binary_logistic.py b/internal/xgjson/testdata/test_binary_logistic.py
new file mode 100644
index 0000000..dad9bd4
--- /dev/null
+++ b/internal/xgjson/testdata/test_binary_logistic.py
@@ -0,0 +1,74 @@
+# /// script
+# dependencies = ["xgboost>=2.0", "numpy", "scikit-learn"]
+# ///
+"""
+Generates binary:logistic XGBoost test model.
+Run with: uv run test_binary_logistic.py
+Outputs: test_binary_logistic.json, test_binary_logistic.ubj, test_binary_logistic_expected.json
+"""
+
+import json
+import os
+import numpy as np
+from sklearn.datasets import make_classification
+import xgboost as xgb
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+
+
+def main():
+    X, y = make_classification(
+        n_samples=1000,
+        n_features=3,
+        n_informative=3,
+        n_redundant=0,
+        n_clusters_per_class=1,
+        random_state=42,
+    )
+
+    X_train, X_test = X[:800], X[800:]
+    y_train = y[:800]
+
+    model = xgb.XGBClassifier(
+        n_estimators=10,
+        max_depth=3,
+        objective="binary:logistic",
+        random_state=42,
+        eval_metric="logloss",
+    )
+    model.fit(X_train, y_train)
+
+    json_path = os.path.join(SCRIPT_DIR, "test_binary_logistic.json")
+    ubj_path = os.path.join(SCRIPT_DIR, "test_binary_logistic.ubj")
+    model.get_booster().save_model(json_path)
+    model.get_booster().save_model(ubj_path)
+    print(f"Saved {json_path}")
+    print(f"Saved {ubj_path}")
+
+    test_rows = X_test[:5]
+    dtest = xgb.DMatrix(test_rows)
+    raw_preds = model.get_booster().predict(dtest, output_margin=True)
+    prob_preds = model.predict_proba(test_rows)[:, 1]
+
+    expected = []
+    for i in range(len(test_rows)):
+        expected.append(
+            {
+                "features": test_rows[i].tolist(),
+                "raw_score": float(raw_preds[i]),
+                "probability": float(prob_preds[i]),
+            }
+        )
+
+    expected_path = os.path.join(SCRIPT_DIR, "test_binary_logistic_expected.json")
+    with open(expected_path, "w") as f:
+        json.dump(expected, f, indent=2)
+    print(f"Saved {expected_path}")
+
+    print(f"\nXGBoost version: {xgb.__version__}")
+    for e in expected:
+        print(f"  features={[round(v,4) for v in e['features']]}  raw={e['raw_score']:.6f}  prob={e['probability']:.6f}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/internal/xgjson/testdata/test_binary_logistic.ubj b/internal/xgjson/testdata/test_binary_logistic.ubj
new file mode 100644
index 0000000..2123e14
Binary files /dev/null and b/internal/xgjson/testdata/test_binary_logistic.ubj differ
diff --git a/internal/xgjson/testdata/test_binary_logistic_expected.json b/internal/xgjson/testdata/test_binary_logistic_expected.json
new file mode 100644
index 0000000..d116578
--- /dev/null
+++ b/internal/xgjson/testdata/test_binary_logistic_expected.json
@@ -0,0 +1,47 @@
+[
+  {
+    "features": [
+      -1.1096492747839761,
+      0.40340213424485294,
+      0.07954668421267086
+    ],
+    "raw_score": -0.7728347778320312,
+    "probability": 0.3158662021160126
+  },
+  {
+    "features": [
+      -1.403825230749223,
+      1.3655851426765793,
+      0.5469370969348255
+    ],
+    "raw_score": 1.4077504873275757,
+    "probability": 0.8034109473228455
+  },
+  {
+    "features": [
+      -0.8717338325160207,
+      1.0131613603890717,
+      1.5004776410785952
+    ],
+    "raw_score": 2.404935121536255,
+    "probability": 0.9172028303146362
+  },
+  {
+    "features": [
+      -0.7083348159161926,
+      1.087972292415291,
+      1.2560764067197803
+    ],
+    "raw_score": 3.208322286605835,
+    "probability": 0.9611462950706482
+  },
+  {
+    "features": [
+      -1.2069009394127748,
+      1.2786601324536067,
+      -0.011722010905850233
+    ],
+    "raw_score": -0.7728347778320312,
+    "probability": 0.3158662021160126
+  }
+]
\ No newline at end of file
diff --git a/internal/xgjson/testdata/test_multiclass.json b/internal/xgjson/testdata/test_multiclass.json
new file mode 100644
index 0000000..356a9f3
--- /dev/null
+++ b/internal/xgjson/testdata/test_multiclass.json
@@ -0,0 +1 @@
+{"learner":{"attributes":{},"feature_names":[],"feature_types":[],"gradient_booster":{"model":{"cats":{"enc":[],"feature_segments":[],"sorted_idx":[]},"gbtree_model_param":{"num_parallel_tree":"1","num_trees":"30"},"iteration_indptr":[0,3,6,9,12,15,18,21,24,27,30],"tree_info":[0,1,2,0,1,2,0,1,2,0,1,2,0,1,2,0,1,2,0,1,2,0,1,2,0,1,2,0,1,2],"trees":[{"base_weights":[8.976922E-8,1.1891924E0,-3.6806875E-1,-1.565994E-1,1.2743015E0,-2.0320473E-2,-5.801027E-1,4.2051739E-1,9.365237E-2,-1.289252E-1,6.1291154E-2,-1.9667643E-1,-6.719352E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":0,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[9.881445E1,8.436264E0,1.2699818E1,0E0,5.8297577E0,6.1886563E0,2.7311745E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-5.058216E-1,-2.000158E0,6.0872287E-1,-1.565994E-1,1.4570175E0,-1.0269129E0,5.924044E-1,4.2051739E-1,9.365237E-2,-1.289252E-1,6.1291154E-2,-1.9667643E-1,-6.719352E-2],"split_indices":[3,0,3,0,2,1,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.23756E2,5.2358906E1,1.713971E2,2.23756E0,5.0121346E1,6.5336754E1,1.0606034E2,4.3856174E1,6.265168E0,2.2823112E1,4.251364E1,8.681733E1,1.9243015E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-1.6837087E-7,4.479209E-1,-5.045543E-1,5.4555234E-2,1.2178032E0,8.804552E-1,-6.205394E-1,-1.6066408E-1,6.819169E-2,3.925975E-1,-1.4185265E-1,-5.911352E-2,3.8677734E-1,-2.1556853E-1,-8.2272686E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":1,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[5.010972E1,3.5503872E1,1.72964E1,8.15283E0,6.7124557E0,3.98522E0,3.088787E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-3.717547E-1,-4.0281177E-1,-1.014656E0,-4.4667578E-1,1.850578E0,-3.530614E-1,1.1128337E0,-1.6066408E-1,6.819169E-2,3.925975E-1,-1.4185265E-1,-5.911352E-2,3.8677734E-1,-2.1556853E-1,-8.2272686E-2],"split_indices":[2,0,0,3,3,3,3,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.1972401E2,1.1645373E2,1.0327029E2,7.77823E1,3.8671425E1,7.4706163E0,9.579967E1,1.7138474E1,6.064383E1,3.6913635E1,1.7577921E0,2.19724E0,5.2733765E0,7.382727E1,2.19724E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[3.2448177E-8,-6.637632E-1,1.5248057E-1,-7.0995367E-1,1.2725432E-1,-1.4294602E-1,6.348038E-1,-2.201662E-1,-1.0262389E-3,-1.0314276E-1,2.357759E-1,-1.4246517E-1,2.6561716E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":2,"left_children":[1,3,5,7,-1,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[2.2782997E1,2.4143295E0,2.617415E1,6.85936E-1,0E0,2.160259E1,1.9806606E1,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,5,5,6,6],"right_children":[2,4,6,8,-1,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-9.3936366E-1,3.0433664E0,-1.8242843E-1,1.5687288E0,1.2725432E-1,1.5553991E0,-3.7348688E-1,-2.201662E-1,-1.0262389E-3,-1.0314276E-1,2.357759E-1,-1.4246517E-1,2.6561716E-1],"split_indices":[3,0,2,1,0,3,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.2310402E2,4.105114E1,1.8205287E2,3.9712513E1,1.3386241E0,1.1333684E2,6.8716034E1,3.837389E1,1.3386241E0,9.370369E1,1.9633154E1,1.2493825E1,5.622221E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-2.0901388E-3,7.066297E-1,-3.7673816E-1,-2.7283943E-1,7.987066E-1,-1.166203E-1,-5.032613E-1,1.8392135E-1,-1.826072E-1,2.6300237E-1,2.2548191E-2,-2.0539433E-1,1.1224851E-2,-1.7371142E-1,-5.1528372E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":3,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[5.8382786E1,7.0756645E0,4.6751633E0,2.5390375E0,3.869522E0,4.2579136E0,2.3238049E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-1.947818E-1,-2.2962296E0,6.0872287E-1,-1.5332314E0,1.5481571E0,-1.6530755E0,5.924044E-1,1.8392135E-1,-1.826072E-1,2.6300237E-1,2.2548191E-2,-2.0539433E-1,1.1224851E-2,-1.7371142E-1,-5.1528372E-2],"split_indices":[3,1,3,3,2,0,2,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.1787793E2,7.503966E1,1.4283827E2,6.2793884E0,6.876027E1,4.73708E1,9.546747E1,1.4992453E0,4.7801433E0,6.1880505E1,6.879761E0,9.370666E0,3.8000137E1,7.70973E1,1.837017E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-1.3279037E-3,3.171575E-1,-4.6415395E-1,2.3645068E-2,8.152216E-1,5.511289E-1,-5.8502877E-1,1.2800452E-1,-3.8428452E-2,2.995066E-1,1.11927696E-1,-1.4434238E-1,2.6591742E-1,-1.937455E-1,-1.0518843E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":4,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[3.192071E1,1.8734081E1,1.1155401E1,5.053966E0,3.4765015E0,3.8548107E0,9.04459E-1,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-2.877946E-1,-4.0281177E-1,-9.587663E-1,-1.1083441E0,-7.3509055E-1,-3.530614E-1,1.1128337E0,1.2800452E-1,-3.8428452E-2,2.995066E-1,1.11927696E-1,-1.4434238E-1,2.6591742E-1,-1.937455E-1,-1.0518843E-1],"split_indices":[2,0,0,1,2,3,3,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.1455482E2,1.2728244E2,8.7272385E1,8.074586E1,4.6536583E1,8.90614E0,7.836624E1,2.1589642E1,5.915621E1,3.2016865E1,1.4519718E1,2.0767446E0,6.8293953E0,6.083427E1,1.7531973E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[3.388171E-3,-5.8328724E-1,1.2091643E-1,-6.3269967E-1,1.24666505E-1,-1.1077672E-1,4.6495634E-1,-1.9731566E-1,2.7590208E-3,-8.5615814E-2,1.7995414E-1,-1.2866977E-1,1.9218363E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":5,"left_children":[1,3,5,7,-1,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[1.5205626E1,2.0794935E0,1.469448E1,5.7564735E-1,0E0,1.3892177E1,1.1912209E1,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,5,5,6,6],"right_children":[2,4,6,8,-1,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-9.3936366E-1,3.0433664E0,-1.8242843E-1,1.5687288E0,1.24666505E-1,1.5553991E0,-3.7348688E-1,-1.9731566E-1,2.7590208E-3,-8.5615814E-2,1.7995414E-1,-1.2866977E-1,1.9218363E-1],"split_indices":[3,0,2,1,0,3,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.185053E2,3.5806103E1,1.8269919E2,3.446757E1,1.3385327E0,1.0958043E2,7.311876E1,3.3150803E1,1.316767E0,8.844642E1,2.1134008E1,1.1772185E1,6.1346577E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-3.627935E-3,5.643761E-1,-3.2544547E-1,-3.6645105E-1,6.4271986E-1,-9.8873034E-2,-4.4622588E-1,-1.9266257E-1,1.1406814E-1,2.2241534E-1,-5.9353437E-2,-1.7357367E-1,1.7972209E-2,-1.5769641E-1,-3.8165253E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":6,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[3.8057682E1,5.726818E0,3.5925999E0,1.578974E0,5.937004E0,3.6521955E0,2.1223698E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-1.947818E-1,-1.7600034E0,6.0872287E-1,-5.051962E-1,1.5481571E0,-1.543846E0,5.924044E-1,-1.9266257E-1,1.1406814E-1,2.2241534E-1,-5.9353437E-2,-1.7357367E-1,1.7972209E-2,-1.5769641E-1,-3.8165253E-2],"split_indices":[3,0,3,2,2,0,2,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.061954E2,7.430112E1,1.3189429E2,5.482639E0,6.881848E1,4.6494225E1,8.540007E1,4.106113E0,1.3765259E0,6.16948E1,7.123678E0,1.0904812E1,3.5589413E1,6.785685E1,1.7543219E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-2.3265986E-3,2.4472557E-1,-4.2390215E-1,2.2704393E-2,7.0940715E-1,5.565847E-1,-5.341267E-1,3.319532E-2,-1.5480052E-1,-8.584619E-2,2.3229305E-1,2.346827E-1,-1.1742429E-1,-1.847629E-1,-8.020587E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":7,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.145066E1,1.3381815E1,8.521408E0,4.269415E0,2.932415E0,2.056324E0,1.3383408E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-2.2600034E-1,-1.3960569E-1,-1.014656E0,1.850578E0,3.7759027E-1,2.864812E-1,1.1128337E0,3.319532E-2,-1.5480052E-1,-8.584619E-2,2.3229305E-1,2.346827E-1,-1.1742429E-1,-1.847629E-1,-8.020587E-2],"split_indices":[2,0,0,3,3,1,3,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.0396097E2,1.2886365E2,7.509732E1,8.790987E1,4.095378E1,7.179908E0,6.791741E1,7.625536E1,1.1654502E1,2.2978382E0,3.865594E1,5.9404693E0,1.2394388E0,5.0985043E1,1.693237E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[5.8407024E-3,-7.454842E-2,6.61053E-1,-4.675233E-1,1.1418001E-1,-1.7490596E-1,9.3166107E-1,-1.788098E-1,1.9935963E-1,5.713473E-2,-2.0596418E-1,9.640786E-2,2.8833634E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":8,"left_children":[1,3,5,7,9,-1,11,-1,-1,-1,-1,-1,-1],"loss_changes":[1.1128508E1,1.4015454E1,8.316058E0,9.310222E0,7.946671E0,0E0,3.9007187E-2,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,6,6],"right_children":[2,4,6,8,10,-1,12,-1,-1,-1,-1,-1,-1],"split_conditions":[1.5481571E0,5.406891E-2,-7.3064186E-2,1.5989536E0,1.9810429E0,-1.7490596E-1,6.802459E-1,-1.788098E-1,1.9935963E-1,5.713473E-2,-2.0596418E-1,9.640786E-2,2.8833634E-1],"split_indices":[2,3,0,1,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[2.0934401E2,1.8725552E2,2.208848E1,6.027411E1,1.2698142E2,3.740002E0,1.8348478E1,5.4554634E1,5.7194715E0,1.1663419E2,1.0347224E1,1.3932775E0,1.6955202E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-6.520784E-3,5.034979E-1,-2.6448688E-1,-3.0338088E-1,5.7349336E-1,-6.451221E-2,-3.969988E-1,-1.7097509E-1,9.132237E-2,1.9818838E-1,-1.3702532E-2,-1.8452558E-1,1.6406905E-2,-1.3624011E-1,1.7421275E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":9,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.5519024E1,3.8538609E0,3.3872843E0,1.121987E0,3.283474E0,3.4807763E0,2.03539E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-2.662504E-1,-1.7600034E0,6.0872287E-1,-5.051962E-1,1.4570175E0,-1.6530755E0,1.8039737E0,-1.7097509E-1,9.132237E-2,1.9818838E-1,-1.3702532E-2,-1.8452558E-1,1.6406905E-2,-1.3624011E-1,1.7421275E-2],"split_indices":[3,0,3,2,2,0,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.9194925E2,6.4156044E1,1.27793205E2,4.855022E0,5.930102E1,5.152433E1,7.6268875E1,3.4182942E0,1.4367279E0,5.1926525E1,7.374498E0,8.429569E0,4.309476E1,6.7763145E1,8.505735E0],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-1.0645691E-3,2.2012648E-1,-3.5846508E-1,1.4714547E-2,5.941625E-1,4.735961E-1,-4.6156082E-1,1.12558745E-1,-3.77464E-2,1.9603299E-1,-1.4754216E-1,1.9787346E-1,-1.1305776E-1,-1.7235164E-1,-5.2932955E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":10,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.5373497E1,9.220861E0,6.595866E0,4.027528E0,3.061223E0,1.6229713E0,2.0450363E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-3.717547E-1,-4.0281177E-1,-1.014656E0,-1.0269129E0,1.850578E0,2.864812E-1,1.1128337E0,1.12558745E-1,-3.77464E-2,1.9603299E-1,-1.4754216E-1,1.9787346E-1,-1.1305776E-1,-1.7235164E-1,-5.2932955E-2],"split_indices":[2,0,0,1,3,1,3,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.9247069E2,1.191279E2,7.334279E1,7.756844E1,4.155945E1,7.6894355E0,6.565335E1,2.1289669E1,5.6278778E1,3.9785908E1,1.7735406E0,6.493782E0,1.1956534E0,4.6298138E1,1.9355217E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[7.2899703E-3,-4.8482883E-1,8.1380315E-2,-5.448387E-1,1.0623165E-1,1.1662075E-2,6.649444E-1,-1.7231289E-1,-4.9774544E-3,-2.6839187E-2,1.4042485E-1,-1.4367083E-1,2.4241816E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":11,"left_children":[1,3,5,7,-1,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[7.3624496E0,1.5232654E0,7.129156E0,3.814006E-1,0E0,7.321403E0,3.4486837E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,5,5,6,6],"right_children":[2,4,6,8,-1,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-9.3936366E-1,1.5989536E0,1.553888E0,2.996951E0,1.0623165E-1,1.7482485E0,-2.538463E-1,-1.7231289E-1,-4.9774544E-3,-2.6839187E-2,1.4042485E-1,-1.4367083E-1,2.4241816E-1],"split_indices":[3,1,2,0,0,3,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.9984319E2,2.5424725E1,1.7441847E2,2.405463E1,1.3700948E0,1.5671567E2,1.7702793E1,2.2698114E1,1.3565165E0,1.2894925E2,2.776642E1,1.7077156E0,1.5995077E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-9.485977E-3,5.173073E-1,-2.0132104E-1,-1.4228354E-1,5.712821E-1,-3.0152557E-2,-3.5412034E-1,1.8488286E-1,-9.377627E-2,-1.0566302E-1,3.632692E-2,-1.3348909E-1,-1.315392E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":12,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[1.8166344E1,2.83951E0,3.4349093E0,0E0,2.0207405E0,3.122415E0,1.9397707E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-5.058216E-1,-2.000158E0,6.0872287E-1,-1.4228354E-1,1.5989536E0,-1.0269129E0,5.924044E-1,1.8488286E-1,-9.377627E-2,-1.0566302E-1,3.632692E-2,-1.3348909E-1,-1.315392E-2],"split_indices":[3,0,3,0,1,1,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.7773206E2,4.6991978E1,1.3074008E2,2.023871E0,4.4968105E1,6.222824E1,6.851183E1,4.307794E1,1.8901649E0,1.9460339E1,4.2767902E1,5.265979E1,1.5852043E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[5.2180095E-4,1.7266046E-1,-3.6015958E-1,1.3660084E-2,5.300831E-1,3.773546E-1,-4.7569916E-1,-1.1902709E-2,2.6027045E-1,1.8343304E-1,-3.5189666E-2,-1.1239097E-1,1.705419E-1,-5.4319464E-2,-1.6800156E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":13,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.1327617E1,7.0113926E0,5.2022257E0,3.9741178E0,2.0731316E0,1.4404948E0,1.175416E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-2.2600034E-1,-1.3960569E-1,-9.587663E-1,2.6605442E0,2.8335938E0,-3.530614E-1,-1.4435455E0,-1.1902709E-2,2.6027045E-1,1.8343304E-1,-3.5189666E-2,-1.1239097E-1,1.705419E-1,-5.4319464E-2,-1.6800156E-1],"split_indices":[2,0,0,1,0,3,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.8044528E2,1.2250154E2,5.7943745E1,8.550337E1,3.6998173E1,7.54112E0,5.0402626E1,8.137421E1,4.129162E0,3.2916832E1,4.0813394E0,1.3335973E0,6.207523E0,1.19137945E1,3.848883E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[8.39162E-3,-4.4714794E-1,6.7530744E-2,-5.1339555E-1,9.307784E-2,7.640994E-3,5.6538594E-1,-1.6428992E-1,1.9248568E-3,-2.700297E-2,1.01534545E-1,-1.4207219E-1,2.1525909E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":14,"left_children":[1,3,5,7,-1,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[5.1692576E0,1.265688E0,5.0557585E0,3.7108374E-1,0E0,4.9567223E0,3.2181735E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,5,5,6,6],"right_children":[2,4,6,8,-1,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-9.3936366E-1,1.5989536E0,1.5481571E0,2.996951E0,9.307784E-2,1.5553991E0,-2.538463E-1,-1.6428992E-1,1.9248568E-3,-2.700297E-2,1.01534545E-1,-1.4207219E-1,2.1525909E-1],"split_indices":[3,1,2,0,0,3,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.8975803E2,2.105025E1,1.6870778E2,1.9653406E1,1.3968422E0,1.5149854E2,1.7209246E1,1.8388708E1,1.2646985E0,1.1753218E2,3.396636E1,1.9283444E0,1.5280901E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-1.2626532E-2,3.045353E-1,-2.5842E-1,-3.556971E-1,4.0415043E-1,-2.9166606E-1,7.693084E-1,-1.5010996E-1,1.1268311E-1,1.4982644E-1,-7.5805195E-2,-1.18902445E-1,-1.5447201E-2,5.9276357E-2,2.8622952E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":15,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.2976739E1,4.9250126E0,3.4277816E0,1.2525285E0,4.108309E0,2.2809978E0,1.5090215E-1,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[5.406891E-2,-1.5891352E0,2.9755492E0,-2.877946E-1,1.5481571E0,4.6791342E-1,1.7804539E-1,-1.5010996E-1,1.1268311E-1,1.4982644E-1,-7.5805195E-2,-1.18902445E-1,-1.5447201E-2,5.9276357E-2,2.8622952E-1],"split_indices":[3,0,0,2,2,2,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.644522E2,7.169758E1,9.275461E1,9.062452E0,6.2635124E1,9.054187E1,2.212743E0,7.8294077E0,1.2330447E0,5.491028E1,7.724848E0,6.2606094E1,2.7935774E1,1.0459479E0,1.1667951E0],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[2.145788E-3,1.712665E-1,-2.877501E-1,1.1636712E-2,4.8477724E-1,4.2308635E-1,-3.870616E-1,9.962344E-2,-3.300727E-2,1.6604947E-1,-1.4108709E-1,1.7400391E-1,-9.964812E-2,-1.6987023E-1,-4.1386925E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":16,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[8.399696E0,5.4090395E0,4.6201315E0,2.8713953E0,2.6545057E0,1.1664158E0,2.417324E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-4.4955203E-1,-4.8287207E-1,-1.014656E0,-1.0269129E0,1.850578E0,2.864812E-1,8.159176E-1,9.962344E-2,-3.300727E-2,1.6604947E-1,-1.4108709E-1,1.7400391E-1,-9.964812E-2,-1.6987023E-1,-4.1386925E-2],"split_indices":[2,0,0,1,3,1,3,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.6932138E2,1.0719482E2,6.212656E1,7.171627E1,3.5478546E1,7.2160983E0,5.491046E1,1.9259457E1,5.245681E1,3.348932E1,1.98923E0,6.166775E0,1.0493231E0,3.1197876E1,2.3712585E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[9.509712E-3,-4.0005688E-2,4.2434493E-1,-3.4024462E-1,7.0136264E-2,-1.5706638E-1,6.43454E-1,-1.5161529E-1,1.8823531E-1,3.856378E-2,-1.911911E-1,6.0481496E-2,2.002758E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":17,"left_children":[1,3,5,7,9,-1,11,-1,-1,-1,-1,-1,-1],"loss_changes":[3.7399385E0,5.4086227E0,4.368891E0,7.328385E0,5.0388503E0,0E0,3.8328648E-2,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,6,6],"right_children":[2,4,6,8,10,-1,12,-1,-1,-1,-1,-1,-1],"split_conditions":[1.5481571E0,5.406891E-2,-7.3064186E-2,1.5989536E0,1.9810429E0,-1.5706638E-1,6.802459E-1,-1.5161529E-1,1.8823531E-1,3.856378E-2,-1.911911E-1,6.0481496E-2,2.002758E-1],"split_indices":[2,3,0,1,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.802478E2,1.6183453E2,1.8413269E1,4.2873997E1,1.1896053E2,3.1962652E0,1.5217004E1,3.703212E1,5.841878E0,1.1064356E2,8.316973E0,1.2724842E0,1.3944519E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-1.3786456E-2,4.232589E-1,-1.5453258E-1,-1.23297E-1,4.7504672E-1,-5.4358995E-1,-9.4504744E-2,1.7179418E-1,1.6227977E-2,-1.7619376E-1,3.070284E-3,1.2520383E-1,-5.3179897E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":18,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[9.444844E0,1.8115249E0,2.6593513E0,0E0,1.4352713E0,3.7659693E-1,4.384926E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-5.058216E-1,-2.000158E0,-1.7600034E0,-1.23297E-1,9.893304E-1,1.7568178E-1,-1.8138796E0,1.7179418E-1,1.6227977E-2,-1.7619376E-1,3.070284E-3,1.2520383E-1,-5.3179897E-2],"split_indices":[3,0,0,0,1,2,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.514803E2,3.6410633E1,1.1506967E2,1.7227559E0,3.4687878E1,1.4304238E1,1.00765434E2,2.7866507E1,6.8213706E0,1.3201552E1,1.1026858E0,1.3462419E1,8.730302E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[3.3527268E-3,-4.1702497E-1,9.393785E-2,1.7970294E-1,-5.137885E-1,2.1195391E-1,-1.9290258E-1,7.9478994E-2,-1.691505E-1,8.768038E-2,-1.2975918E-1,8.9961015E-2,-1.0973162E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":19,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[6.1304564E0,3.1138854E0,4.5234766E0,0E0,1.1859779E0,5.0072107E0,3.442437E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-3.4058717E-1,-2.850353E0,-3.717547E-1,1.7970294E-1,-2.399855E0,1.850578E0,-4.6335888E-1,7.9478994E-2,-1.691505E-1,8.768038E-2,-1.2975918E-1,8.9961015E-2,-1.0973162E-1],"split_indices":[3,1,2,0,0,3,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.5896027E2,2.7542112E1,1.3141815E2,1.9435924E0,2.559852E1,9.329461E1,3.812355E1,1.2865309E0,2.4311989E1,8.34427E1,9.851909E0,9.71002E0,2.841353E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[9.11443E-3,-3.864425E-1,4.776072E-2,-4.6021813E-1,8.523351E-2,-6.5666504E-2,2.1095058E-1,-1.5276419E-1,1.9150585E-2,-7.7280626E-2,1.0416328E-1,1.0260255E-1,-1.8611878E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":20,"left_children":[1,3,5,7,-1,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[2.6447253E0,8.926511E-1,2.9273286E0,3.9855504E-1,0E0,7.5084167E0,7.2701354E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,5,5,6,6],"right_children":[2,4,6,8,-1,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-9.3936366E-1,3.0433664E0,-1.8242843E-1,1.5687288E0,8.523351E-2,1.0844262E0,1.6783428E0,-1.5276419E-1,1.9150585E-2,-7.7280626E-2,1.0416328E-1,1.0260255E-1,-1.8611878E-1],"split_indices":[3,0,2,1,0,1,1,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.7079967E2,1.4401037E1,1.5639862E2,1.3256236E1,1.1448007E0,9.261971E1,6.377892E1,1.2148722E1,1.1075149E0,6.348386E1,2.9135849E1,5.5602116E1,8.176804E0],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-1.2654852E-2,4.538828E-1,-1.0541267E-1,4.9121177E-1,-2.8421942E-2,-5.156002E-1,-4.9706228E-2,-5.619883E-2,1.627176E-1,-1.6806856E-1,5.9875115E-3,1.1171465E-1,-3.6554433E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":21,"left_children":[1,3,5,7,-1,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[6.18774E0,5.322919E-1,2.697228E0,8.78458E-1,0E0,3.4917474E-1,3.2737217E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,5,5,6,6],"right_children":[2,4,6,8,-1,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-9.3936366E-1,1.5989536E0,-1.7600034E0,-1.6238737E0,-2.8421942E-2,1.7568178E-1,-1.8138796E0,-5.619883E-2,1.627176E-1,-1.6806856E-1,5.9875115E-3,1.1171465E-1,-3.6554433E-2],"split_indices":[3,1,0,0,0,2,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.4088081E2,2.2719069E1,1.1816175E2,2.1369694E1,1.3493742E0,1.3141353E1,1.050204E2,1.3122215E0,2.0057472E1,1.2087991E1,1.053362E0,1.4722142E1,9.0298256E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[3.821175E-3,1.24513425E-1,-2.3297627E-1,-6.062304E-3,4.2964745E-1,3.8214183E-1,-3.2982954E-1,-1.7670315E-2,2.2356662E-1,1.4584626E-1,-1.2079639E-1,-4.772599E-2,1.683034E-1,-1.6292854E-1,-2.9046834E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":22,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[4.3606954E0,4.0286374E0,3.195314E0,2.8921854E0,1.620605E0,8.125682E-1,2.1858974E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-4.4955203E-1,-4.0281177E-1,-1.014656E0,2.6605442E0,1.850578E0,-3.530614E-1,8.159176E-1,-1.7670315E-2,2.2356662E-1,1.4584626E-1,-1.2079639E-1,-4.772599E-2,1.683034E-1,-1.6292854E-1,-2.9046834E-2],"split_indices":[2,0,0,1,3,3,3,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.505656E2,1.0004718E2,5.0518425E1,7.075093E1,2.9296247E1,6.4715886E0,4.4046837E1,6.6962944E1,3.787986E0,2.7823166E1,1.4730812E0,1.6349894E0,4.8365993E0,2.2302801E1,2.1744038E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[7.387525E-3,-3.6329985E-2,3.1401834E-1,3.922867E-2,-4.8384723E-1,-4.8578045E-1,1.9148001E-1,-3.420363E-2,7.310421E-2,-1.8171448E-1,1.0716105E-1,-2.429568E-2,-1.6776676E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":23,"left_children":[1,3,5,7,9,11,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.218499E0,4.921865E0,5.664933E0,3.9433513E0,2.3708324E0,1.4866972E-1,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5],"right_children":[2,4,6,8,10,12,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[1.4425881E0,1.3956678E0,4.5153925E-1,1.0111414E0,1.2216467E0,-5.167576E-1,1.9148001E-1,-3.420363E-2,7.310421E-2,-1.8171448E-1,1.0716105E-1,-2.429568E-2,-1.6776676E-1],"split_indices":[2,0,0,3,2,1,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.6364516E2,1.4399655E2,1.9648613E1,1.2397671E2,2.0019842E1,5.5242333E0,1.4124379E1,7.1116135E1,5.286057E1,1.7730515E1,2.2893276E0,1.1712577E0,4.3529754E0],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-1.4467765E-2,2.080743E-1,-1.7292127E-1,-4.7704002E-1,2.9902416E-1,-2.0685807E-1,6.884434E-1,-7.605224E-3,-1.697991E-1,1.3227023E-1,-1.7589947E-2,-8.156897E-2,6.693411E-2,5.9669532E-2,2.4860035E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":24,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[4.736835E0,3.6256235E0,2.4345884E0,2.643479E-1,2.5610108E0,2.191857E0,5.507636E-2,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[5.406891E-2,-2.2962296E0,2.9755492E0,-1.2172549E0,9.893304E-1,1.8039737E0,1.7804539E-1,-7.605224E-3,-1.697991E-1,1.3227023E-1,-1.7589947E-2,-8.156897E-2,6.693411E-2,5.9669532E-2,2.4860035E-1],"split_indices":[3,1,0,3,1,1,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.3231004E2,5.489662E1,7.741342E1,5.9360414E0,4.896058E1,7.5210075E1,2.2033439E0,1.1881337E0,4.7479076E0,3.4888256E1,1.4072324E1,6.56477E1,9.562378E0,1.0290837E0,1.1742603E0],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[6.7778416E-3,-3.662384E-1,7.065992E-2,1.2101906E-1,-5.0399894E-1,3.104683E-1,-2.9183773E-2,-4.156303E-3,-1.6096625E-1,1.2665306E-1,-9.17492E-2,2.9581118E-2,-1.1535417E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":25,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[3.4348352E0,2.4382627E0,2.953601E0,0E0,2.8495455E-1,2.6238682E0,4.0220222E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-3.4058717E-1,-2.2962296E0,-8.278937E-1,1.2101906E-1,-2.600786E0,1.8622017E0,-5.490728E-1,-4.156303E-3,-1.6096625E-1,1.2665306E-1,-9.17492E-2,2.9581118E-2,-1.1535417E-1],"split_indices":[3,1,1,0,2,3,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.4205885E2,2.0079405E1,1.2197945E2,2.755703E0,1.7323702E1,3.5236805E1,8.674264E1,1.1677582E0,1.6155943E1,3.009657E1,5.1402345E0,6.4330086E1,2.2412552E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[6.1122016E-3,-3.0932928E-2,2.8709438E-1,3.4882322E-2,-4.5499435E-1,-4.5712167E-1,1.8004636E-1,-6.688488E-2,4.297451E-2,-1.7304227E-1,9.31752E-2,-1.6844256E-2,-1.6021039E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":26,"left_children":[1,3,5,7,9,11,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.6421589E0,3.9102547E0,4.555378E0,3.4327443E0,1.9340532E0,1.5531588E-1,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5],"right_children":[2,4,6,8,10,12,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[1.4425881E0,1.3956678E0,4.5153925E-1,-8.182202E-1,1.2216467E0,-5.167576E-1,1.8004636E-1,-6.688488E-2,4.297451E-2,-1.7304227E-1,9.31752E-2,-1.6844256E-2,-1.6021039E-1],"split_indices":[2,0,0,1,2,1,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.5590987E2,1.385351E2,1.7374771E1,1.20717255E2,1.7817837E1,5.008424E0,1.2366348E1,3.5409992E1,8.530727E1,1.5584688E1,2.233149E0,1.0845095E0,3.9239142E0],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[-1.2841339E-2,3.8976517E-1,-7.814242E-2,-2.4121646E-2,4.3709022E-1,-4.9403644E-1,-2.5205292E-2,1.4776048E-1,-4.215324E-2,-1.6158266E-1,-4.3613804E-3,1.0649363E-1,-2.8236344E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0],"id":27,"left_children":[1,3,5,-1,7,9,11,-1,-1,-1,-1,-1,-1],"loss_changes":[3.3149943E0,4.2681456E-1,2.3725522E0,0E0,5.8634734E-1,2.524433E-1,2.5788443E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,4,4,5,5,6,6],"right_children":[2,4,6,-1,8,10,12,-1,-1,-1,-1,-1,-1],"split_conditions":[-9.3936366E-1,-1.842745E0,-1.7600034E0,-2.4121646E-2,1.5365523E0,-5.051962E-1,-1.8138796E0,1.4776048E-1,-4.215324E-2,-1.6158266E-1,-4.3613804E-3,1.0649363E-1,-2.8236344E-2],"split_indices":[3,0,0,0,1,2,2,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2393182E2,1.660246E1,1.0732936E2,1.4543734E0,1.51480875E1,1.117803E1,9.615133E1,1.3955733E1,1.1923542E0,1.0114544E1,1.063486E0,1.4117768E1,8.203356E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"13","size_leaf_vector":"1"}},{"base_weights":[5.4487037E-3,1.499641E-1,-1.4004897E-1,-7.177455E-2,6.129975E-1,5.26862E-1,-2.9116732E-1,-4.130257E-2,1.8700607E-1,2.0334715E-1,-1.04372114E-1,-6.4618096E-2,2.0535553E-1,-1.618584E-1,-5.0461195E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":28,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.9090364E0,7.168543E0,7.10337E0,2.2762642E0,1.6420355E0,1.6798639E0,1.6458578E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-9.239956E-1,-7.7296084E-1,-1.0505564E0,2.6605442E0,1.850578E0,-3.530614E-1,5.4416007E-1,-4.130257E-2,1.8700607E-1,2.0334715E-1,-1.04372114E-1,-6.4618096E-2,2.0535553E-1,-1.618584E-1,-5.0461195E-2],"split_indices":[2,0,0,1,3,3,3,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.363511E2,6.839106E1,6.796004E1,4.681646E1,2.15746E1,1.2095245E1,5.5864796E1,4.3494926E1,3.3215344E0,2.0487062E1,1.0875387E0,2.0548177E0,1.0040428E1,1.73777E1,3.8487095E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[5.6611793E-3,3.773538E-2,-2.841661E-1,-3.950626E-2,3.0515108E-1,3.110922E-1,-5.7242596E-1,-4.867336E-2,1.1077137E-1,2.9061523E-1,-5.970359E-2,2.6886097E-1,-1.14180885E-1,-2.0935065E-1,7.355219E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":29,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.4161686E0,2.842019E0,2.7830777E0,5.437348E0,1.0503294E1,2.6079679E0,1.2559624E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[1.976097E0,1.2041476E0,1.2879819E-1,7.043773E-1,-4.8287207E-1,-1.1052026E0,5.9545773E-1,-4.867336E-2,1.1077137E-1,2.9061523E-1,-5.970359E-2,2.6886097E-1,-1.14180885E-1,-2.0935065E-1,7.355219E-2],"split_indices":[1,1,3,2,0,2,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.5018854E2,1.3600693E2,1.41816225E1,1.06187485E2,2.9819439E1,4.601105E0,9.580518E0,8.21288E1,2.4058685E1,1.247661E1,1.7342829E1,2.3324249E0,2.2686806E0,8.43342E0,1.1470972E0],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}}]},"name":"gbtree"},"learner_model_param":{"base_score":"[1.4027715E-2,-2.2120595E-2,8.092999E-3]","boost_from_average":"1","num_class":"3","num_feature":"4","num_target":"1"},"objective":{"name":"multi:softprob","softmax_multiclass_param":{"num_class":"3"}}},"version":[3,2,0]}
\ No newline at end of file
diff --git a/internal/xgjson/testdata/test_multiclass.py b/internal/xgjson/testdata/test_multiclass.py
new file mode 100644
index 0000000..169184d
--- /dev/null
+++ b/internal/xgjson/testdata/test_multiclass.py
@@ -0,0 +1,97 @@
+# /// script
+# dependencies = ["xgboost>=2.0", "numpy", "scikit-learn"]
+# ///
+"""
+Generates multi:softprob XGBoost test model (3-class).
+Run with: uv run test_multiclass.py
+Outputs: test_multiclass.json, test_multiclass.ubj, test_multiclass_expected.json
+
+NOTE on base_score: XGBoost 3.x stores a per-class base_score vector for multi-class models.
+The leaves Go library (xgEnsemble) only supports a single scalar BaseScore, so it uses 0.0
+for multi-class. Expected values here are the raw tree sums (output_margin minus per-class
+base_score) and the softmax of those sums, which is what the Go library actually computes.
+"""
+
+import json
+import os
+import numpy as np
+from sklearn.datasets import make_classification
+import xgboost as xgb
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+
+
+def softmax(x):
+    e = np.exp(x - x.max(axis=1, keepdims=True))
+    return e / e.sum(axis=1, keepdims=True)
+
+
+def main():
+    X, y = make_classification(
+        n_samples=600,
+        n_features=4,
+        n_classes=3,
+        n_informative=4,
+        n_redundant=0,
+        random_state=42,
+    )
+
+    X_train, X_test = X[:500], X[500:]
+    y_train = y[:500]
+
+    model = xgb.XGBClassifier(
+        n_estimators=10,
+        max_depth=3,
+        objective="multi:softprob",
+        num_class=3,
+        random_state=42,
+    )
+    model.fit(X_train, y_train)
+
+    json_path = os.path.join(SCRIPT_DIR, "test_multiclass.json")
+    ubj_path = os.path.join(SCRIPT_DIR, "test_multiclass.ubj")
+    model.get_booster().save_model(json_path)
+    model.get_booster().save_model(ubj_path)
+    print(f"Saved {json_path}")
+    print(f"Saved {ubj_path}")
+
+    # Parse per-class base_scores from the saved model.
+    # XGBoost 3.x stores these as "[v0,v1,...,vN]" in learner_model_param.base_score.
+    with open(json_path) as f:
+        saved = json.load(f)
+    bs_str = saved["learner"]["learner_model_param"]["base_score"].strip("[]")
+    base_scores = np.array([float(x) for x in bs_str.split(",")])
+    print(f"Per-class base_scores: {base_scores.tolist()}")
+
+    test_rows = X_test[:5]
+    dtest = xgb.DMatrix(test_rows)
+    # output_margin=True gives per-class raw scores including per-class base_score
+    raw_preds = model.get_booster().predict(dtest, output_margin=True)  # shape: [n, 3]
+
+    # Go uses BaseScore=0 for multi-class (xgEnsemble doesn't support per-class base_score).
+    # Subtract per-class base_scores to get the pure tree sums that Go computes.
+    go_raw_preds = raw_preds - base_scores  # broadcast over rows
+    go_probs = softmax(go_raw_preds)
+
+    expected = []
+    for i in range(len(test_rows)):
+        expected.append(
+            {
+                "features": test_rows[i].tolist(),
+                "raw_scores": go_raw_preds[i].tolist(),
+                "probabilities": go_probs[i].tolist(),
+            }
+        )
+
+    expected_path = os.path.join(SCRIPT_DIR, "test_multiclass_expected.json")
+    with open(expected_path, "w") as f:
+        json.dump(expected, f, indent=2)
+    print(f"Saved {expected_path}")
+
+    print(f"\nXGBoost version: {xgb.__version__}")
+    for e in expected:
+        print(f"  raw={[round(v,4) for v in e['raw_scores']]}  prob={[round(v,4) for v in e['probabilities']]}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/internal/xgjson/testdata/test_multiclass.ubj b/internal/xgjson/testdata/test_multiclass.ubj
new file mode 100644
index 0000000..4a2c619
Binary files /dev/null and b/internal/xgjson/testdata/test_multiclass.ubj differ
diff --git a/internal/xgjson/testdata/test_multiclass_expected.json b/internal/xgjson/testdata/test_multiclass_expected.json
new file mode 100644
index 0000000..69766d5
--- /dev/null
+++ b/internal/xgjson/testdata/test_multiclass_expected.json
@@ -0,0 +1,92 @@
+[
+  {
+    "features": [
+      1.2898756580524573,
+      0.9248578543855488,
+      -1.2090147492408292,
+      1.1638837750179802
+    ],
+    "raw_scores": [
+      -1.1162556412542726,
+      1.9363673925616456,
+      -0.1567775006479492
+    ],
+    "probabilities": [
+      0.04035327308558963,
+      0.8543112105802694,
+      0.10533551633414107
+    ]
+  },
+  {
+    "features": [
+      0.14181489267063685,
+      -2.4478019338481474,
+      0.812756653818766,
+      1.1106575556359448
+    ],
+    "raw_scores": [
+      -0.5212679508055115,
+      -0.9247214793942261,
+      0.7192503814512023
+    ],
+    "probabilities": [
+      0.1951061845696842,
+      0.13033270078894887,
+      0.674561114641367
+    ]
+  },
+  {
+    "features": [
+      0.08734803051372408,
+      0.9105387201543911,
+      0.7767547146693972,
+      -0.7705641418952007
+    ],
+    "raw_scores": [
+      1.6781063077127074,
+      -1.7592041492245483,
+      0.2956783180051575
+    ],
+    "probabilities": [
+      0.7793506344474285,
+      0.0250569340114191,
+      0.19559243154115238
+    ]
+  },
+  {
+    "features": [
+      1.2445468253247949,
+      1.0397630318427356,
+      -1.1561191254665355,
+      1.0761159326795893
+    ],
+    "raw_scores": [
+      -1.1162556412542726,
+      1.9363673925616456,
+      -0.1567775006479492
+    ],
+    "probabilities": [
+      0.04035327308558963,
+      0.8543112105802694,
+      0.10533551633414107
+    ]
+  },
+  {
+    "features": [
+      -0.32039158134005596,
+      2.703654404944862,
+      -1.9104368617091088,
+      1.0037481848182237
+    ],
+    "raw_scores": [
+      -0.35270831015953064,
+      1.8141070604541016,
+      -0.24331870629255675
+    ],
+    "probabilities": [
+      0.0921996052286053,
+      0.8049428003930158,
+      0.10285759437837882
+    ]
+  }
+]
\ No newline at end of file
diff --git a/internal/xgjson/testdata/test_poisson.json b/internal/xgjson/testdata/test_poisson.json
new file mode 100644
index 0000000..b26e740
--- /dev/null
+++ b/internal/xgjson/testdata/test_poisson.json
@@ -0,0 +1 @@
+{"learner":{"attributes":{},"feature_names":[],"feature_types":[],"gradient_booster":{"model":{"cats":{"enc":[],"feature_segments":[],"sorted_idx":[]},"gbtree_model_param":{"num_parallel_tree":"1","num_trees":"10"},"iteration_indptr":[0,1,2,3,4,5,6,7,8,9,10],"tree_info":[0,0,0,0,0,0,0,0,0,0],"trees":[{"base_weights":[1.608796E-8,-1.6029176E-1,7E-1,-3.0662945E-1,8.970203E-2,2.1499042E-1,7E-1,-1.1313498E-1,-5.6605335E-2,-3.312177E-2,1.3046043E-1,4.193799E-3,1.391802E-1,2.1000001E-1,3.4098327E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":0,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.5571881E2,3.7720955E1,2.4467834E1,5.3337936E0,2.6378124E1,5.30269E0,9.165039E-1,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[8.384892E-1,3.4935063E-1,-2.5862848E-2,6.1704636E-2,2.172312E-1,1.3206617E0,1.8212988E0,-1.1313498E-1,-5.6605335E-2,-3.312177E-2,1.3046043E-1,4.193799E-3,1.391802E-1,2.1000001E-1,3.4098327E-2],"split_indices":[2,0,0,2,2,2,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2364442E3,1.0293398E3,2.071044E2,6.491332E2,3.8020657E2,1.05097755E2,1.02006645E2,4.049355E2,2.4419772E2,2.4110661E2,1.3909998E2,5.87311E1,4.6366657E1,9.8915535E1,3.0911105E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[6.0249087E-3,-1.4485276E-1,6.292016E-1,-2.8846076E-1,7.249133E-2,1.7385581E-1,7E-1,-1.2007688E-1,-6.294832E-2,-2.9236889E-2,9.6497245E-2,-3.8690425E-2,7.703155E-2,2.1000001E-1,2.9422885E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":1,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.1504521E2,3.0768625E1,3.072609E1,5.132305E0,1.6661806E1,2.972922E0,1.0310974E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[8.384892E-1,3.4935063E-1,1.1451292E-2,-4.0526927E-1,2.172312E-1,1.0027578E0,1.8212988E0,-1.2007688E-1,-6.294832E-2,-2.9236889E-2,9.6497245E-2,-3.8690425E-2,7.703155E-2,2.1000001E-1,2.9422885E-2],"split_indices":[2,0,0,2,2,2,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2216106E3,9.841138E2,2.3749681E2,5.9237836E2,3.9173547E2,1.16082E2,1.2141481E2,2.4292001E2,3.494583E2,2.3325153E2,1.5848395E2,2.483281E1,9.124919E1,1.1821648E2,3.1983294E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[8.249043E-3,-1.3017151E-1,4.8842916E-1,-2.6980826E-1,5.9279393E-2,1.3929035E-1,7E-1,-1.0460437E-1,-4.492878E-2,-2.764657E-2,7.164046E-2,7.21004E-2,-1.7957078E-2,1.4741372E-1,2.1000001E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":2,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[8.0968575E1,2.5032322E1,2.6119629E1,5.103985E0,1.0952632E1,2.4984605E0,4.2014847E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[8.384892E-1,3.4935063E-1,1.1451292E-2,6.1704636E-2,1.5704857E-1,2.0279728E-1,1.3591876E0,-1.0460437E-1,-4.492878E-2,-2.764657E-2,7.164046E-2,7.21004E-2,-1.7957078E-2,1.4741372E-1,2.1000001E-1],"split_indices":[2,0,0,2,2,1,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2162241E3,9.446428E2,2.7158136E2,5.435727E2,4.010701E2,1.22446434E2,1.4913492E2,3.268921E2,2.1668063E2,2.1781798E2,1.8325212E2,8.109383E1,4.1352604E1,1.0679396E2,4.234096E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[9.134995E-3,-1.2018245E-1,3.7619865E-1,-2.7463865E-1,2.4330707E-2,2.2345707E-1,7E-1,1.6753457E-2,-8.807992E-2,-8.588739E-2,3.03916E-2,-5.0609536E-2,7.969856E-2,2.1000001E-1,4.92894E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":3,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[5.772286E1,2.0079765E1,2.5428822E1,2.7408333E0,1.115958E1,4.354599E0,1.4625626E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[8.2798815E-1,1.599916E-1,1.3591876E0,-1.5338614E0,-9.5888263E-1,8.978154E-1,4.438127E-1,1.6753457E-2,-8.807992E-2,-8.588739E-2,3.03916E-2,-5.0609536E-2,7.969856E-2,2.1000001E-1,4.92894E-2],"split_indices":[2,0,0,1,2,2,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2140918E3,8.982934E2,3.1579843E2,4.3377383E2,4.6451956E2,2.5939658E2,5.6401833E1,2.3431297E1,4.1034253E2,9.171107E1,3.7280847E2,2.491316E1,2.3448343E2,5.059793E1,5.8039017E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[9.09531E-3,-2.0343041E-1,1.8307982E-1,-2.3020588E-1,4.771158E-1,1.12988435E-1,7E-1,2.3115667E-2,-7.8853436E-2,5.2792482E-2,2.1000001E-1,-1.2131449E-2,6.406971E-2,2.1000001E-1,2.3619696E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":4,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[4.4970226E1,1.0085991E1,2.5175674E1,5.3096256E0,1.9333873E0,9.15984E0,7.2229195E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[2.344009E-1,1.9050602E0,1.5112284E0,-1.3204889E0,2.0931683E0,-5.119711E-2,4.438127E-1,2.3115667E-2,-7.8853436E-2,5.2792482E-2,2.1000001E-1,-1.2131449E-2,6.406971E-2,2.1000001E-1,2.3619696E-2],"split_indices":[2,0,0,1,0,0,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2141871E3,5.464805E2,6.677066E2,5.2643036E2,2.0050182E1,5.9203174E2,7.567487E1,5.042025E1,4.7601007E2,1.1633196E1,8.4169855E0,2.3466206E2,3.5736966E2,5.6691853E1,1.898302E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[8.181038E-3,-8.697243E-2,3.0043316E-1,-2.3265421E-1,2.8890641E-2,2.0781858E-1,7E-1,-1.0578607E-1,-4.8775777E-2,9.829865E-2,-7.159475E-3,1.6448982E-2,8.489892E-2,2.1000001E-1,3.727326E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":5,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[3.3880238E1,1.5521219E1,1.4152052E1,3.3662758E0,8.0914755E0,2.947919E0,1.9109726E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[9.5205367E-1,1.599916E-1,1.5112284E0,-4.0526927E-1,-1.1719716E0,-4.5647115E-1,3.8913205E-2,-1.0578607E-1,-4.8775777E-2,9.829865E-2,-7.159475E-3,1.6448982E-2,8.489892E-2,2.1000001E-1,3.727326E-2],"split_indices":[2,0,0,2,1,0,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2163882E3,9.181522E2,2.9823593E2,4.062903E2,5.1186188E2,2.5665982E2,4.157613E1,1.4831805E2,2.5797226E2,7.603583E1,4.3582605E2,8.513578E1,1.7152403E2,3.5333263E1,6.242867E0],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[7.0725908E-3,-1.7376871E-1,1.3126637E-1,-2.003981E-1,3.637069E-1,9.796992E-2,6.8071556E-1,-8.719458E-2,-2.7215052E-2,3.962268E-2,1.8250549E-1,-2.0084638E-2,4.5219928E-2,1.21616825E-1,2.1000001E-1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":6,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.7424751E1,7.1887817E0,1.3176699E1,4.669132E0,1.2697597E0,5.9593873E0,1.1783447E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[2.344009E-1,1.9050602E0,2.0769806E0,-4.0526927E-1,2.0931683E0,-3.8318732E-1,2.44513E0,-8.719458E-2,-2.7215052E-2,3.962268E-2,1.8250549E-1,-2.0084638E-2,4.5219928E-2,1.21616825E-1,2.1000001E-1],"split_indices":[2,0,0,2,0,0,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2190674E3,4.9617575E2,7.228916E2,4.7335062E2,2.2825144E1,6.826986E2,4.0193024E1,2.5877194E2,2.1457866E2,1.2515371E1,1.0309773E1,1.654118E2,5.172868E2,1.8398088E1,2.1794937E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[5.7086255E-3,-1.5682158E-1,1.0779569E-1,-2.0145491E-1,9.470294E-2,2.705563E-1,4.000411E-2,-8.737209E-2,-2.8699575E-2,4.7959937E-3,1.3134255E-1,2.499932E-2,1.3513407E-1,-2.7228395E-2,2.8268065E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":7,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.0315517E1,5.32216E0,8.286631E0,3.7857876E0,1.9340974E0,7.403927E0,3.7810888E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[2.344009E-1,1.3634287E0,-5.8487415E-1,-4.0526927E-1,2.0931683E0,5.4873884E-1,-7.971821E-2,-8.737209E-2,-2.8699575E-2,4.7959937E-3,1.3134255E-1,2.499932E-2,1.3513407E-1,-2.7228395E-2,2.8268065E-2],"split_indices":[2,0,1,2,0,0,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2223805E3,4.7137704E2,7.510034E2,4.0050586E2,7.0871185E1,2.1994563E2,5.310578E2,2.1569179E2,1.8481407E2,5.8497177E1,1.2374003E1,1.0849266E2,1.1145297E2,1.5546336E2,3.7559442E2],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[3.986343E-3,-6.855007E-2,1.8679157E-1,-1.9220346E-1,1.4377538E-2,3.6981606E-1,1.1181358E-1,3.1648107E-2,-6.5965526E-2,-6.833215E-2,1.5860908E-2,3.80832E-2,1.7194837E-1,1.7563814E-2,5.6632925E-2],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":8,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.6290684E1,9.022352E0,4.7478733E0,2.933526E0,4.928169E0,4.914747E0,1.0097888E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[9.5205367E-1,1.599916E-1,-5.7094365E-1,-1.2964723E0,-9.5888263E-1,-8.6334527E-1,1.5317789E0,3.1648107E-2,-6.5965526E-2,-6.833215E-2,1.5860908E-2,3.80832E-2,1.7194837E-1,1.7563814E-2,5.6632925E-2],"split_indices":[2,0,1,1,2,1,2,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2265896E3,8.785925E2,3.4799716E2,3.5216357E2,5.2642896E2,9.998851E1,2.4800864E2,2.9721193E1,3.2244238E2,7.152883E1,4.549001E2,4.6305855E1,5.368266E1,1.4760732E2,1.0040132E2],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[2.6625937E-3,-1.5288876E-1,6.671782E-2,-1.8247098E-1,2.5947097E-1,1.1514038E-2,1.8340808E-1,8.038846E-2,-6.2043E-2,4.460892E-2,1.2815052E-1,-1.8337945E-2,3.2311738E-2,6.9908835E-2,-2.6620878E-3],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":9,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.2274185E1,4.430501E0,5.618476E0,3.7374134E0,3.9095032E-1,4.1527863E0,2.6750078E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[3.0631492E-2,1.9050602E0,8.705556E-1,-2.0729835E0,-6.9893044E-1,8.978154E-1,4.438127E-1,8.038846E-2,-6.2043E-2,4.460892E-2,1.2815052E-1,-1.8337945E-2,3.2311738E-2,6.9908835E-2,-2.6620878E-3],"split_indices":[2,0,0,1,2,2,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[1.2298446E3,3.5831866E2,8.7152594E2,3.3485413E2,2.3464561E1,5.9238116E2,2.7914478E2,1.6653053E1,3.1820105E2,1.5266671E1,8.19789E0,3.3771536E2,2.5466583E2,2.217156E2,5.7429165E1],"tree_param":{"num_deleted":"0","num_feature":"3","num_nodes":"15","size_leaf_vector":"1"}}]},"name":"gbtree"},"learner_model_param":{"base_score":"[1.535E0]","boost_from_average":"1","num_class":"0","num_feature":"3","num_target":"1"},"objective":{"name":"count:poisson","poisson_regression_param":{"max_delta_step":"0.699999988"}}},"version":[3,2,0]}
\ No newline at end of file
diff --git a/internal/xgjson/testdata/test_poisson.py b/internal/xgjson/testdata/test_poisson.py
new file mode 100644
index 0000000..44d2a1a
--- /dev/null
+++ b/internal/xgjson/testdata/test_poisson.py
@@ -0,0 +1,69 @@
+# /// script
+# dependencies = ["xgboost>=2.0", "numpy", "scikit-learn"]
+# ///
+"""
+Generates count:poisson XGBoost test model.
+Run with: uv run test_poisson.py
+Outputs: test_poisson.json, test_poisson.ubj, test_poisson_expected.json
+"""
+
+import json
+import os
+import numpy as np
+import xgboost as xgb
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+
+
+def main():
+    rng = np.random.default_rng(42)
+    X = rng.standard_normal((500, 3))
+    y = rng.poisson(np.exp(X @ [0.5, -0.3, 0.8]))
+
+    X_train, X_test = X[:400], X[400:]
+    y_train = y[:400]
+
+    model = xgb.XGBRegressor(
+        n_estimators=10,
+        max_depth=3,
+        objective="count:poisson",
+        random_state=42,
+    )
+    model.fit(X_train, y_train)
+
+    json_path = os.path.join(SCRIPT_DIR, "test_poisson.json")
+    ubj_path = os.path.join(SCRIPT_DIR, "test_poisson.ubj")
+    model.get_booster().save_model(json_path)
+    model.get_booster().save_model(ubj_path)
+    print(f"Saved {json_path}")
+    print(f"Saved {ubj_path}")
+
+    test_rows = X_test[:5]
+    dtest = xgb.DMatrix(test_rows)
+    # output_margin=True gives log-space raw scores
+    raw_preds = model.get_booster().predict(dtest, output_margin=True)
+    # predict() gives exp(margin) = count prediction
+    count_preds = model.predict(test_rows)
+
+    expected = []
+    for i in range(len(test_rows)):
+        expected.append(
+            {
+                "features": test_rows[i].tolist(),
+                "raw_score": float(raw_preds[i]),
+                "prediction": float(count_preds[i]),
+            }
+        )
+
+    expected_path = os.path.join(SCRIPT_DIR, "test_poisson_expected.json")
+    with open(expected_path, "w") as f:
+        json.dump(expected, f, indent=2)
+    print(f"Saved {expected_path}")
+
+    print(f"\nXGBoost version: {xgb.__version__}")
+    for e in expected:
+        print(f"  features={[round(v,4) for v in e['features']]}  raw={e['raw_score']:.6f}  pred={e['prediction']:.6f}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/internal/xgjson/testdata/test_poisson.ubj b/internal/xgjson/testdata/test_poisson.ubj
new file mode 100644
index 0000000..c70dae2
Binary files /dev/null and b/internal/xgjson/testdata/test_poisson.ubj differ
diff --git a/internal/xgjson/testdata/test_poisson_expected.json b/internal/xgjson/testdata/test_poisson_expected.json
new file mode 100644
index 0000000..6a23356
--- /dev/null
+++ b/internal/xgjson/testdata/test_poisson_expected.json
@@ -0,0 +1,47 @@
+[
+  {
+    "features": [
+      -1.322541133291679,
+      -0.4861941306990469,
+      0.42022670947458574
+    ],
+    "raw_score": -0.01655573584139347,
+    "prediction": 0.9835805296897888
+  },
+  {
+    "features": [
+      -0.10239670661109637,
+      -0.6505636204055921,
+      -0.674212461724509
+    ],
+    "raw_score": -0.48458048701286316,
+    "prediction": 0.6159555315971375
+  },
+  {
+    "features": [
+      -0.7123370644127087,
+      -0.8795096286589426,
+      2.2816328772760617
+    ],
+    "raw_score": 0.8761682510375977,
+    "prediction": 2.401679515838623
+  },
+  {
+    "features": [
+      0.2975110711855549,
+      0.8867590544047386,
+      -0.4890774681189237
+    ],
+    "raw_score": -0.18565593659877777,
+    "prediction": 0.8305593132972717
+  },
+  {
+    "features": [
+      -0.18595667718604886,
+      -0.713553981370355,
+      -2.651708334919202
+    ],
+    "raw_score": -0.48458048701286316,
+    "prediction": 0.6159555315971375
+  }
+]
\ No newline at end of file
diff --git a/internal/xgjson/testdata/test_regression.json b/internal/xgjson/testdata/test_regression.json
new file mode 100644
index 0000000..825add4
--- /dev/null
+++ b/internal/xgjson/testdata/test_regression.json
@@ -0,0 +1 @@
+{"learner":{"attributes":{},"feature_names":[],"feature_types":[],"gradient_booster":{"model":{"cats":{"enc":[],"feature_segments":[],"sorted_idx":[]},"gbtree_model_param":{"num_parallel_tree":"1","num_trees":"10"},"iteration_indptr":[0,1,2,3,4,5,6,7,8,9,10],"tree_info":[0,0,0,0,0,0,0,0,0,0],"trees":[{"base_weights":[-3.4900674E-7,-2.1703766E1,3.888591E1,-5.660842E1,1.1113851E0,1.8131002E1,6.76406E1,-2.410375E1,-1.3382771E1,-2.6366928E0,1.2017798E1,-1.8681777E0,9.061715E0,1.3893869E1,2.3894E1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":0,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[3.3927622E5,2.055225E5,8.471297E4,2.5468469E4,6.0891074E4,2.5188213E4,1.1675E4,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[3.3198032E-1,-2.0812225E-1,1.9204912E-1,-7.2713715E-1,1.2897527E0,-7.728777E-1,6.982233E-1,-2.410375E1,-1.3382771E1,-2.6366928E0,1.2017798E1,-1.8681777E0,9.061715E0,1.3893869E1,2.3894E1],"split_indices":[1,0,0,1,0,0,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,2.57E2,1.43E2,1.01E2,1.56E2,8.4E1,5.9E1,3.2E1,6.9E1,1.25E2,3.1E1,2.8E1,5.6E1,2.3E1,3.6E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[4.208488E-2,-3.1665228E1,1.6616203E1,-4.209725E1,-9.3845826E-1,-1.0655563E1,3.008117E1,-1.8151125E1,-9.645155E0,-1.0015606E1,2.8573256E0,-7.5468655E0,3.6063957E0,6.482903E0,1.6121021E1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":1,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.1125997E5,4.4195344E4,9.7264875E4,1.69125E4,1.2539539E4,2.9239414E4,3.4309188E4,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-3.357847E-1,6.725737E-1,-2.5497723E-1,-6.095122E-1,-1.3044695E0,6.4537597E-1,8.700677E-1,-1.8151125E1,-9.645155E0,-1.0015606E1,2.8573256E0,-7.5468655E0,3.6063957E0,6.482903E0,1.6121021E1],"split_indices":[1,0,0,0,1,1,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,1.37E2,2.63E2,1.02E2,3.5E1,8.7E1,1.76E2,3.4E1,6.8E1,8E0,2.7E1,5.3E1,3.4E1,1.31E2,4.5E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-1.1167165E-2,-2.2418953E1,1.1702036E1,-5.0537853E1,-1.5123926E1,-1.5157427E0,2.8579166E1,-1.5612584E1,-1.4890947E0,-7.604434E0,1.3045177E-1,-1.0831369E1,9.6371937E-1,5.392886E0,1.2125985E1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":2,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.0551161E5,2.754336E4,5.893595E4,1.6610234E3,1.7668422E4,2.4483562E4,1.3836781E4,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-3.357847E-1,-1.3719012E0,1.9204912E-1,1.2897527E0,2.5049284E-1,-1.5255252E0,8.004095E-1,-1.5612584E1,-1.4890947E0,-7.604434E0,1.3045177E-1,-1.0831369E1,9.6371937E-1,5.392886E0,1.2125985E1],"split_indices":[1,1,0,0,0,0,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,1.37E2,2.63E2,2.7E1,1.1E2,1.48E2,1.15E2,2.6E1,1E0,6.6E1,4.4E1,1.7E1,1.31E2,6.2E1,5.3E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-5.002281E-2,-7.8147173E0,2.0562166E1,-2.491751E1,-3.1489772E-1,1.1856391E0,2.5955784E1,-9.092211E0,4.703793E-1,-4.403968E0,1.952447E0,-5.087349E0,2.3423016E0,5.6828566E0,1.2984971E1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":3,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[6.4339684E4,3.7446367E4,1.1465273E4,1.2752523E4,2.009022E4,3.1360671E3,9.630484E3,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[5.5979043E-1,-4.6341768E-1,-7.539646E-1,9.1786194E-1,-6.414816E-1,-5.100164E-1,1.3993554E0,-9.092211E0,4.703793E-1,-4.403968E0,1.952447E0,-5.087349E0,2.3423016E0,5.6828566E0,1.2984971E1],"split_indices":[1,0,0,3,1,3,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,2.91E2,1.09E2,8.8E1,2.03E2,2.4E1,8.5E1,7.3E1,1.5E1,6.5E1,1.38E2,6E0,1.8E1,6.2E1,2.3E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-2.2922823E-2,-8.812916E0,9.98283E0,-1.2287054E1,2.057175E1,3.8232608E0,2.4433542E1,-9.349065E0,-2.6622648E0,-7.8267634E-1,7.2537894E0,2.2643733E-1,8.955402E0,5.292737E0,1.1337043E1],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":4,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[3.535613E4,2.209925E4,1.6640496E4,1.2092553E4,1.9864639E3,1.0599696E4,4.4128164E3,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[1.042011E-1,1.5033983E0,5.6091946E-1,-9.983854E-1,-1.2239403E0,1.6777008E0,1.1581109E0,-9.349065E0,-2.6622648E0,-7.8267634E-1,7.2537894E0,2.2643733E-1,8.955402E0,5.292737E0,1.1337043E1],"split_indices":[1,0,0,0,2,1,0,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,2.13E2,1.87E2,1.91E2,2.2E1,1.32E2,5.5E1,2.8E1,1.63E2,3E0,1.9E1,1.19E2,1.3E1,3.8E1,1.7E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[3.450642E-2,-7.6010866E0,7.9013066E0,-2.0408184E1,-2.7366667E0,-1.424733E1,1.0195322E1,-1.4414116E1,-5.0269175E0,-2.0192778E0,4.2070208E0,-5.7176957E0,-9.828203E-1,-1.6186364E0,4.080827E0],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":5,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[2.41472E4,1.2653163E4,1.0205502E4,4.8470625E3,1.0069055E4,9.405698E2,9.635811E3,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[5.9218433E-2,-5.449191E-1,-1.2608839E0,-2.1356742E0,8.696059E-1,1.943843E-1,-8.254972E-1,-1.4414116E1,-5.0269175E0,-2.0192778E0,4.2070208E0,-5.7176957E0,-9.828203E-1,-1.6186364E0,4.080827E0],"split_indices":[3,1,0,1,0,1,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,2.03E2,1.97E2,5.5E1,1.48E2,1.8E1,1.79E2,5E0,5E1,1.2E2,2.8E1,1.2E1,6E0,3.2E1,1.47E2],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-2.9440423E-2,-5.389619E0,6.9236445E0,-7.6967645E0,1.5408404E1,-1.2216392E1,8.663874E0,-4.9256887E0,-1.1134292E0,-1.1573362E0,5.547367E0,-5.031951E0,1.3524102E0,1.9145297E0,6.08035E0],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":6,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.4982496E4,1.10109795E4,5.93475E3,7.065041E3,1.4374014E3,1.2187563E3,4.134176E3,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[1.504189E-1,1.3902075E0,-1.2515395E0,-5.100164E-1,-1.51937E0,1.3071427E0,9.633761E-1,-4.9256887E0,-1.1134292E0,-1.1573362E0,5.547367E0,-5.031951E0,1.3524102E0,1.9145297E0,6.08035E0],"split_indices":[0,1,1,3,0,2,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,2.26E2,1.74E2,2.04E2,2.2E1,1.4E1,1.6E2,6.3E1,1.41E2,3E0,1.9E1,1.1E1,3E0,1.35E2,2.5E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-1.395906E-2,-2.2662601E1,1.1718702E0,-2.7021698E1,-7.9877605E0,-1.3520322E0,1.1025952E1,-9.467001E0,-4.487939E0,-3.1854148E0,7.7455217E-1,-1.2673126E0,2.162385E0,1.5491321E0,4.4347887E0],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":7,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[1.0796384E4,1.0635391E3,9.515531E3,3.4842383E2,1.9422064E2,7.529547E3,1.6414043E3,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[-1.5566292E0,5.8392817E-1,9.1786194E-1,-1.7131345E0,1.521316E0,6.725737E-1,-3.426876E-1,-9.467001E0,-4.487939E0,-3.1854148E0,7.7455217E-1,-1.2673126E0,2.162385E0,1.5491321E0,4.4347887E0],"split_indices":[1,0,3,1,2,0,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,1.9E1,3.81E2,1.4E1,5E0,3.04E2,7.7E1,9E0,5E0,4E0,1E0,2.28E2,7.6E1,3.1E1,4.6E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-3.3429414E-2,-2.614427E0,7.8663254E0,-1.4629532E1,-7.3140913E-1,-1.6345493E1,9.057828E0,-6.7656684E0,-3.291841E0,-1.1315507E0,1.3278493E0,-5.9210744E0,-4.1697222E-1,1.190232E0,3.8547845E0],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":8,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[8.196654E3,6.8445713E3,3.0040518E3,9.4035547E2,4.140856E3,2.261709E2,1.7825186E3,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[6.454842E-1,-1.0676204E0,-1.6685841E0,-8.0227727E-1,3.8240975E-1,1.0666747E0,-1.0594835E-1,-6.7656684E0,-3.291841E0,-1.1315507E0,1.3278493E0,-5.9210744E0,-4.1697222E-1,1.190232E0,3.8547845E0],"split_indices":[1,0,3,3,3,1,2,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,3.02E2,9.8E1,4E1,2.62E2,4E0,9.4E1,1.1E1,2.9E1,1.65E2,9.7E1,3E0,1E0,4.1E1,5.3E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}},{"base_weights":[-5.1216763E-2,-9.203662E-1,1.65521E1,-4.6210365E0,2.68967E0,1.094314E0,1.9215572E1,-4.8870697E0,-9.8923075E-1,-6.7744327E-1,1.7080133E0,-1.0934024E0,1.1666605E0,2.6789975E0,6.5080333E0],"categories":[],"categories_nodes":[],"categories_segments":[],"categories_sizes":[],"default_left":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"id":9,"left_children":[1,3,5,7,9,11,13,-1,-1,-1,-1,-1,-1,-1,-1],"loss_changes":[5.8019707E3,5.1157793E3,8.023999E2,2.8654568E3,2.8960432E3,6.71471E1,2.3954541E2,0E0,0E0,0E0,0E0,0E0,0E0,0E0,0E0],"parents":[2147483647,0,0,1,1,2,2,3,3,4,4,5,5,6,6],"right_children":[2,4,6,8,10,12,14,-1,-1,-1,-1,-1,-1,-1,-1],"split_conditions":[1.6420152E0,4.886007E-2,-1.279577E0,-1.5625459E0,-2.8865865E-1,-1.0372461E0,-5.642476E-1,-4.8870697E0,-9.8923075E-1,-6.7744327E-1,1.7080133E0,-1.0934024E0,1.1666605E0,2.6789975E0,6.5080333E0],"split_indices":[0,1,2,1,2,1,1,0,0,0,0,0,0,0,0],"split_type":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"sum_hessian":[4E2,3.81E2,1.9E1,1.88E2,1.93E2,3E0,1.6E1,1.8E1,1.7E2,7.3E1,1.2E2,1E0,2E0,4E0,1.2E1],"tree_param":{"num_deleted":"0","num_feature":"4","num_nodes":"15","size_leaf_vector":"1"}}]},"name":"gbtree"},"learner_model_param":{"base_score":"[3.5165477E0]","boost_from_average":"1","num_class":"0","num_feature":"4","num_target":"1"},"objective":{"name":"reg:squarederror","reg_loss_param":{"scale_pos_weight":"1"}}},"version":[3,2,0]}
\ No newline at end of file
diff --git a/internal/xgjson/testdata/test_regression.py b/internal/xgjson/testdata/test_regression.py
new file mode 100644
index 0000000..5a4629f
--- /dev/null
+++ b/internal/xgjson/testdata/test_regression.py
@@ -0,0 +1,65 @@
+# /// script
+# dependencies = ["xgboost>=2.0", "numpy", "scikit-learn"]
+# ///
+"""
+Generates reg:squarederror XGBoost test model.
+Run with: uv run test_regression.py
+Outputs: test_regression.json, test_regression.ubj, test_regression_expected.json
+"""
+
+import json
+import os
+import numpy as np
+from sklearn.datasets import make_regression
+import xgboost as xgb
+
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+
+
+def main():
+    X, y = make_regression(n_samples=500, n_features=4, random_state=42)
+
+    X_train, X_test = X[:400], X[400:]
+    y_train = y[:400]
+
+    model = xgb.XGBRegressor(
+        n_estimators=10,
+        max_depth=3,
+        objective="reg:squarederror",
+        random_state=42,
+    )
+    model.fit(X_train, y_train)
+
+    json_path = os.path.join(SCRIPT_DIR, "test_regression.json")
+    ubj_path = os.path.join(SCRIPT_DIR, "test_regression.ubj")
+    model.get_booster().save_model(json_path)
+    model.get_booster().save_model(ubj_path)
+    print(f"Saved {json_path}")
+    print(f"Saved {ubj_path}")
+
+    test_rows = X_test[:5]
+    dtest = xgb.DMatrix(test_rows)
+    # identity link: output_margin == predict()
+    raw_preds = model.get_booster().predict(dtest, output_margin=True)
+
+    expected = []
+    for i in range(len(test_rows)):
+        expected.append(
+            {
+                "features": test_rows[i].tolist(),
+                "raw_score": float(raw_preds[i]),
+            }
+        )
+
+    expected_path = os.path.join(SCRIPT_DIR, "test_regression_expected.json")
+    with open(expected_path, "w") as f:
+        json.dump(expected, f, indent=2)
+    print(f"Saved {expected_path}")
+
+    print(f"\nXGBoost version: {xgb.__version__}")
+    for e in expected:
+        print(f"  features={[round(v,4) for v in e['features']]}  raw={e['raw_score']:.6f}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/internal/xgjson/testdata/test_regression.ubj b/internal/xgjson/testdata/test_regression.ubj
new file mode 100644
index 0000000..59db117
Binary files /dev/null and b/internal/xgjson/testdata/test_regression.ubj differ
diff --git a/internal/xgjson/testdata/test_regression_expected.json b/internal/xgjson/testdata/test_regression_expected.json
new file mode 100644
index 0000000..2a50522
--- /dev/null
+++ b/internal/xgjson/testdata/test_regression_expected.json
@@ -0,0 +1,47 @@
+[
+  {
+    "features": [
+      -2.123895724309807,
+      -0.8397218421807761,
+      -0.5993926454440222,
+      -0.525755021680761
+    ],
+    "raw_score": -80.28502655029297
+  },
+  {
+    "features": [
+      0.7326400772155792,
+      0.7276295436369798,
+      0.05194588580729943,
+      -0.08071658010858232
+    ],
+    "raw_score": 57.882362365722656
+  },
+  {
+    "features": [
+      2.4457519796168263,
+      0.2799686263198203,
+      -1.1254890472983765,
+      0.1292211819752275
+    ],
+    "raw_score": 60.96694564819336
+  },
+  {
+    "features": [
+      -0.3653215513121087,
+      2.013387247526623,
+      0.13653533108273744,
+      0.1846803058649084
+    ],
+    "raw_score": 53.012428283691406
+  },
+  {
+    "features": [
+      0.19652116970147013,
+      0.6427227598675439,
+      1.3291525301324314,
+      0.7090037575885123
+    ],
+    "raw_score": 42.95940399169922
+  }
+]
\ No newline at end of file
diff --git a/internal/xgjson/testdata/train.py b/internal/xgjson/testdata/train.py
new file mode 100644
index 0000000..fc8cecc
--- /dev/null
+++ b/internal/xgjson/testdata/train.py
@@ -0,0 +1,21 @@
+# /// script
+# dependencies = []
+# ///
+"""
+Coordinator: runs all per-case training scripts.
+Run with: uv run train.py
+"""
+
+import subprocess
+import pathlib
+
+here = pathlib.Path(__file__).parent
+
+for s in [
+    "test_binary_logistic.py",
+    "test_regression.py",
+    "test_multiclass.py",
+    "test_poisson.py",
+]:
+    print(f"\n=== Running {s} ===")
+    subprocess.run(["uv", "run", str(here / s)], check=True)
diff --git a/internal/xgjson/xgjson_io.go b/internal/xgjson/xgjson_io.go
new file mode 100644
index 0000000..df3f5cc
--- /dev/null
+++ b/internal/xgjson/xgjson_io.go
@@ -0,0 +1,468 @@
+// Package xgjson reads XGBoost models saved in JSON (.json) and Universal Binary JSON (.ubj) formats.
+// These are the modern formats produced by XGBoost 2.x and 3.x.
+//
+// Use ReadModelJSON for .json files and ReadModelUBJ for .ubj files.
+// Both return *ModelJSON with the same logical structure.
+package xgjson
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"strconv"
+	"strings"
+
+	"github.com/dmitryikh/leaves/internal/ubjdecode"
+)
+
+// ModelJSON is the top-level XGBoost model structure.
+type ModelJSON struct {
+	Learner LearnerJSON
+	Version []int32
+}
+
+// LearnerJSON holds the learner sub-tree of the model.
+type LearnerJSON struct {
+	FeatureNames      []string
+	GradientBooster   GradientBoosterJSON
+	LearnerModelParam LearnerModelParamJSON
+	Objective         ObjectiveJSON
+}
+
+// LearnerModelParamJSON holds global model parameters.
+// Numeric fields are stored as strings (or bracketed strings like "[5.2E-1]") by XGBoost.
+type LearnerModelParamJSON struct {
+	BaseScore  string // may be "[value]" in XGBoost 2.x
+	NumClass   string
+	NumFeature string
+}
+
+// ObjectiveJSON holds the objective function name.
+type ObjectiveJSON struct {
+	Name string
+}
+
+// GradientBoosterJSON wraps the booster model and its name.
+type GradientBoosterJSON struct {
+	Model GBTreeModelDataJSON
+	Name  string // "gbtree" or "dart"
+}
+
+// GBTreeModelDataJSON holds the tree ensemble data.
+type GBTreeModelDataJSON struct {
+	GBTreeModelParam GBTreeModelParamJSON
+	TreeInfo         []int32
+	Trees            []TreeJSON
+}
+
+// GBTreeModelParamJSON holds ensemble-level parameters.
+type GBTreeModelParamJSON struct {
+	NumParallelTree string
+	NumTrees        string
+}
+
+// TreeJSON holds a single decision tree's data.
+type TreeJSON struct {
+	ID              int32
+	BaseWeights     []float64
+	DefaultLeft     []int32
+	LeftChildren    []int32
+	RightChildren   []int32
+	SplitConditions []float64
+	SplitIndices    []int32
+	SplitType       []int32
+	NumNodes        string // from tree_param.num_nodes
+}
+
+// ---------------------------------------------------------------------------
+// JSON format (encoding/json with struct tags)
+// ---------------------------------------------------------------------------
+
+// jsonModel is the raw JSON-tagged equivalent used only for unmarshaling.
+type jsonModel struct {
+	Learner struct {
+		FeatureNames    []string `json:"feature_names"`
+		GradientBooster struct {
+			Model struct {
+				GBTreeModelParam struct {
+					NumParallelTree string `json:"num_parallel_tree"`
+					NumTrees        string `json:"num_trees"`
+				} `json:"gbtree_model_param"`
+				TreeInfo []int32 `json:"tree_info"`
+				Trees    []struct {
+					ID              int32     `json:"id"`
+					BaseWeights     []float64 `json:"base_weights"`
+					DefaultLeft     []int32   `json:"default_left"`
+					LeftChildren    []int32   `json:"left_children"`
+					RightChildren   []int32   `json:"right_children"`
+					SplitConditions []float64 `json:"split_conditions"`
+					SplitIndices    []int32   `json:"split_indices"`
+					SplitType       []int32   `json:"split_type"`
+					TreeParam       struct {
+						NumNodes string `json:"num_nodes"`
+					} `json:"tree_param"`
+				} `json:"trees"`
+			} `json:"model"`
+			Name string `json:"name"`
+		} `json:"gradient_booster"`
+		LearnerModelParam struct {
+			BaseScore  string `json:"base_score"`
+			NumClass   string `json:"num_class"`
+			NumFeature string `json:"num_feature"`
+		} `json:"learner_model_param"`
+		Objective struct {
+			Name string `json:"name"`
+		} `json:"objective"`
+	} `json:"learner"`
+	Version []int32 `json:"version"`
+}
+
+// ReadModelJSON reads an XGBoost model from a JSON reader.
+func ReadModelJSON(r io.Reader) (*ModelJSON, error) {
+	var raw jsonModel
+	if err := json.NewDecoder(r).Decode(&raw); err != nil {
+		return nil, fmt.Errorf("xgjson: JSON decode: %w", err)
+	}
+	return jsonModelToModelJSON(&raw), nil
+}
+
+func jsonModelToModelJSON(raw *jsonModel) *ModelJSON {
+	m := &ModelJSON{
+		Version: raw.Version,
+	}
+	rl := &raw.Learner
+	m.Learner.FeatureNames = rl.FeatureNames
+	m.Learner.LearnerModelParam = LearnerModelParamJSON{
+		BaseScore:  rl.LearnerModelParam.BaseScore,
+		NumClass:   rl.LearnerModelParam.NumClass,
+		NumFeature: rl.LearnerModelParam.NumFeature,
+	}
+	m.Learner.Objective = ObjectiveJSON{Name: rl.Objective.Name}
+
+	gb := &rl.GradientBooster
+	m.Learner.GradientBooster = GradientBoosterJSON{
+		Name: gb.Name,
+		Model: GBTreeModelDataJSON{
+			GBTreeModelParam: GBTreeModelParamJSON{
+				NumParallelTree: gb.Model.GBTreeModelParam.NumParallelTree,
+				NumTrees:        gb.Model.GBTreeModelParam.NumTrees,
+			},
+			TreeInfo: gb.Model.TreeInfo,
+		},
+	}
+
+	for _, t := range gb.Model.Trees {
+		m.Learner.GradientBooster.Model.Trees = append(
+			m.Learner.GradientBooster.Model.Trees,
+			TreeJSON{
+				ID:              t.ID,
+				BaseWeights:     t.BaseWeights,
+				DefaultLeft:     t.DefaultLeft,
+				LeftChildren:    t.LeftChildren,
+				RightChildren:   t.RightChildren,
+				SplitConditions: t.SplitConditions,
+				SplitIndices:    t.SplitIndices,
+				SplitType:       t.SplitType,
+				NumNodes:        t.TreeParam.NumNodes,
+			},
+		)
+	}
+	return m
+}
+
+// ---------------------------------------------------------------------------
+// UBJ format — walk map[string]interface{} from ubjdecode
+// ---------------------------------------------------------------------------
+
+// ReadModelUBJ reads an XGBoost model from a UBJ reader.
+// Uses ubjdecode internally; produces the same ModelJSON struct.
+func ReadModelUBJ(r io.Reader) (*ModelJSON, error) {
+	raw, err := ubjdecode.DecodeValue(r)
+	if err != nil {
+		return nil, fmt.Errorf("xgjson: UBJ decode: %w", err)
+	}
+	root, ok := raw.(map[string]interface{})
+	if !ok {
+		return nil, fmt.Errorf("xgjson: UBJ root is not an object")
+	}
+	return ubjMapToModelJSON(root)
+}
+
+// ---------------------------------------------------------------------------
+// UBJ map-walking helpers
+// ---------------------------------------------------------------------------
+
+func ubjMapToModelJSON(root map[string]interface{}) (*ModelJSON, error) {
+	m := &ModelJSON{}
+
+	if v, ok := root["version"]; ok {
+		m.Version = toInt32Slice(v)
+	}
+
+	learnerRaw, ok := root["learner"]
+	if !ok {
+		return nil, fmt.Errorf("xgjson: UBJ missing 'learner' key")
+	}
+	learnerMap, ok := learnerRaw.(map[string]interface{})
+	if !ok {
+		return nil, fmt.Errorf("xgjson: UBJ 'learner' is not an object")
+	}
+
+	// feature_names
+	if v, ok := learnerMap["feature_names"]; ok {
+		m.Learner.FeatureNames = toStringSlice(v)
+	}
+
+	// learner_model_param
+	if v, ok := learnerMap["learner_model_param"]; ok {
+		if mp, ok := v.(map[string]interface{}); ok {
+			m.Learner.LearnerModelParam = LearnerModelParamJSON{
+				BaseScore:  mapString(mp, "base_score"),
+				NumClass:   mapString(mp, "num_class"),
+				NumFeature: mapString(mp, "num_feature"),
+			}
+		}
+	}
+
+	// objective
+	if v, ok := learnerMap["objective"]; ok {
+		if om, ok := v.(map[string]interface{}); ok {
+			m.Learner.Objective = ObjectiveJSON{Name: mapString(om, "name")}
+		}
+	}
+
+	// gradient_booster
+	if v, ok := learnerMap["gradient_booster"]; ok {
+		gbMap, ok := v.(map[string]interface{})
+		if !ok {
+			return nil, fmt.Errorf("xgjson: UBJ 'gradient_booster' is not an object")
+		}
+		gb, err := ubjGradientBooster(gbMap)
+		if err != nil {
+			return nil, err
+		}
+		m.Learner.GradientBooster = *gb
+	}
+
+	return m, nil
+}
+
+func ubjGradientBooster(gbMap map[string]interface{}) (*GradientBoosterJSON, error) {
+	gb := &GradientBoosterJSON{
+		Name: mapString(gbMap, "name"),
+	}
+
+	modelRaw, ok := gbMap["model"]
+	if !ok {
+		return nil, fmt.Errorf("xgjson: UBJ gradient_booster missing 'model'")
+	}
+	modelMap, ok := modelRaw.(map[string]interface{})
+	if !ok {
+		return nil, fmt.Errorf("xgjson: UBJ gradient_booster 'model' is not an object")
+	}
+
+	// gbtree_model_param
+	if v, ok := modelMap["gbtree_model_param"]; ok {
+		if pm, ok := v.(map[string]interface{}); ok {
+			gb.Model.GBTreeModelParam = GBTreeModelParamJSON{
+				NumParallelTree: mapString(pm, "num_parallel_tree"),
+				NumTrees:        mapString(pm, "num_trees"),
+			}
+		}
+	}
+
+	// tree_info
+	if v, ok := modelMap["tree_info"]; ok {
+		gb.Model.TreeInfo = toInt32Slice(v)
+	}
+
+	// trees
+	treesRaw, ok := modelMap["trees"]
+	if !ok {
+		return nil, fmt.Errorf("xgjson: UBJ model missing 'trees'")
+	}
+	treesSlice, ok := treesRaw.([]interface{})
+	if !ok {
+		return nil, fmt.Errorf("xgjson: UBJ 'trees' is not an array")
+	}
+	for i, tRaw := range treesSlice {
+		tMap, ok := tRaw.(map[string]interface{})
+		if !ok {
+			return nil, fmt.Errorf("xgjson: UBJ tree %d is not an object", i)
+		}
+		tree, err := ubjTree(tMap)
+		if err != nil {
+			return nil, fmt.Errorf("xgjson: UBJ tree %d: %w", i, err)
+		}
+		gb.Model.Trees = append(gb.Model.Trees, *tree)
+	}
+
+	return gb, nil
+}
+
+func ubjTree(tMap map[string]interface{}) (*TreeJSON, error) {
+	t := &TreeJSON{}
+
+	if v, ok := tMap["id"]; ok {
+		t.ID = int32(toInt64(v))
+	}
+	t.BaseWeights = toFloat64Slice(tMap["base_weights"])
+	t.DefaultLeft = toInt32Slice(tMap["default_left"])
+	t.LeftChildren = toInt32Slice(tMap["left_children"])
+	t.RightChildren = toInt32Slice(tMap["right_children"])
+	t.SplitConditions = toFloat64Slice(tMap["split_conditions"])
+	t.SplitIndices = toInt32Slice(tMap["split_indices"])
+	t.SplitType = toInt32Slice(tMap["split_type"])
+
+	if v, ok := tMap["tree_param"]; ok {
+		if pm, ok := v.(map[string]interface{}); ok {
+			t.NumNodes = mapString(pm, "num_nodes")
+		}
+	}
+	return t, nil
+}
+
+// ---------------------------------------------------------------------------
+// Type coercion helpers
+// ---------------------------------------------------------------------------
+
+func mapString(m map[string]interface{}, key string) string {
+	if v, ok := m[key]; ok {
+		if s, ok := v.(string); ok {
+			return s
+		}
+		return fmt.Sprintf("%v", v)
+	}
+	return ""
+}
+
+func toInt64(v interface{}) int64 {
+	switch x := v.(type) {
+	case int64:
+		return x
+	case int32:
+		return int64(x)
+	case float64:
+		return int64(x)
+	case float32:
+		return int64(x)
+	default:
+		return 0
+	}
+}
+
+// toInt32Slice coerces any of the UBJ typed array forms to []int32.
+func toInt32Slice(v interface{}) []int32 {
+	if v == nil {
+		return nil
+	}
+	switch x := v.(type) {
+	case []int32:
+		return x
+	case []int64:
+		out := make([]int32, len(x))
+		for i, vv := range x {
+			out[i] = int32(vv)
+		}
+		return out
+	case []interface{}:
+		out := make([]int32, len(x))
+		for i, vv := range x {
+			out[i] = int32(toInt64(vv))
+		}
+		return out
+	default:
+		return nil
+	}
+}
+
+// toFloat64Slice coerces any of the UBJ typed array forms to []float64.
+func toFloat64Slice(v interface{}) []float64 {
+	if v == nil {
+		return nil
+	}
+	switch x := v.(type) {
+	case []float64:
+		return x
+	case []float32:
+		out := make([]float64, len(x))
+		for i, vv := range x {
+			out[i] = float64(vv)
+		}
+		return out
+	case []interface{}:
+		out := make([]float64, len(x))
+		for i, vv := range x {
+			switch f := vv.(type) {
+			case float64:
+				out[i] = f
+			case float32:
+				out[i] = float64(f)
+			case int64:
+				out[i] = float64(f)
+			}
+		}
+		return out
+	default:
+		return nil
+	}
+}
+
+func toStringSlice(v interface{}) []string {
+	if v == nil {
+		return nil
+	}
+	switch x := v.(type) {
+	case []string:
+		return x
+	case []interface{}:
+		out := make([]string, len(x))
+		for i, vv := range x {
+			if s, ok := vv.(string); ok {
+				out[i] = s
+			} else {
+				out[i] = fmt.Sprintf("%v", vv)
+			}
+		}
+		return out
+	default:
+		return nil
+	}
+}
+
+// ParseBaseScore parses XGBoost's base_score string, which may be wrapped in
+// brackets like "[5.2313304E-1]" (XGBoost 2.x) or a plain float string.
+// For multi-class models (XGBoost 3.x) with comma-separated per-class values,
+// use ParseBaseScoreMulti instead.
+func ParseBaseScore(s string) (float64, error) {
+	s = strings.TrimSpace(s)
+	s = strings.TrimPrefix(s, "[")
+	s = strings.TrimSuffix(s, "]")
+	s = strings.TrimSpace(s)
+	v, err := strconv.ParseFloat(s, 64)
+	if err != nil {
+		return 0, fmt.Errorf("xgjson: parse base_score %q: %w", s, err)
+	}
+	return v, nil
+}
+
+// ParseBaseScoreMulti parses XGBoost's base_score string, returning one float64
+// per output group. Handles both single-value (all groups share same score) and
+// comma-separated per-group strings like "[1.4E-2,-2.2E-2,8.1E-3]" from XGBoost 3.x
+// multi-class models.
+func ParseBaseScoreMulti(s string) ([]float64, error) {
+	s = strings.TrimSpace(s)
+	s = strings.TrimPrefix(s, "[")
+	s = strings.TrimSuffix(s, "]")
+	s = strings.TrimSpace(s)
+	parts := strings.Split(s, ",")
+	out := make([]float64, len(parts))
+	for i, p := range parts {
+		v, err := strconv.ParseFloat(strings.TrimSpace(p), 64)
+		if err != nil {
+			return nil, fmt.Errorf("xgjson: parse base_score[%d] in %q: %w", i, s, err)
+		}
+		out[i] = v
+	}
+	return out, nil
+}
diff --git a/xgensemble_io.go b/xgensemble_io.go
index 5102dfc..59d68b5 100644
--- a/xgensemble_io.go
+++ b/xgensemble_io.go
@@ -4,8 +4,11 @@ import (
 	"bufio"
 	"fmt"
 	"os"
+	"path/filepath"
+	"strings"
 
 	"github.com/dmitryikh/leaves/internal/xgbin"
+	"github.com/dmitryikh/leaves/internal/xgjson"
 	"github.com/dmitryikh/leaves/transformation"
 )
 
@@ -248,7 +251,24 @@ func XGEnsembleFromReader(reader *bufio.Reader, loadTransformation bool) (*Ensem
 	return &Ensemble{e, transform}, nil
 }
 
-// XGEnsembleFromFile reads XGBoost model from binary file. Works with 'gbtree' and 'dart' models
+// XGEnsembleFromFile reads an XGBoost model from a file.
+//
+// Format detection uses both the file extension and a content heuristic so
+// that misnamed files don't silently produce wrong results and to preserve
+// backward compatibility with existing callers:
+//
+//   - .json        → content-verified JSON; falls through to legacy binary if
+//                    content doesn't look like XGBoost JSON
+//   - .ubj/.ubjson → content-verified UBJ; falls through to legacy binary if
+//                    content doesn't look like XGBoost UBJ
+//   - any other    → falls through directly to legacy binary reader
+//
+// Note: for unknown extensions, we intentionally do NOT run content detection
+// and fall straight through to the legacy binary reader. This preserves the
+// existing behavior for all callers who weren't using .json/.ubj extensions.
+// The trade-off is that a JSON/UBJ file named e.g. "model.bin" won't be
+// auto-detected; callers should use XGEnsembleFromJSONFile / XGEnsembleFromUBJFile
+// in that case.
 func XGEnsembleFromFile(filename string, loadTransformation bool) (*Ensemble, error) {
 	reader, err := os.Open(filename)
 	if err != nil {
@@ -256,5 +276,21 @@ func XGEnsembleFromFile(filename string, loadTransformation bool) (*Ensemble, er
 	}
 	defer reader.Close()
 	bufReader := bufio.NewReader(reader)
+	ext := strings.ToLower(filepath.Ext(filename))
+	switch ext {
+	case ".json":
+		if ok, _ := xgjson.LooksLikeJSON(bufReader); ok {
+			return XGEnsembleFromJSONReader(bufReader, loadTransformation)
+		}
+		// Extension says JSON but content doesn't match — fall through to legacy
+		// binary. TODO: consider returning an error here instead, since a caller
+		// who named the file .json almost certainly intended it to be JSON.
+	case ".ubj", ".ubjson":
+		if ok, _ := xgjson.LooksLikeUBJ(bufReader); ok {
+			return XGEnsembleFromUBJReader(bufReader, loadTransformation)
+		}
+		// Extension says UBJ but content doesn't match — fall through to legacy binary.
+		// TODO: consider returning an error here instead.
+	}
 	return XGEnsembleFromReader(bufReader, loadTransformation)
 }
diff --git a/xgensemble_json_io.go b/xgensemble_json_io.go
new file mode 100644
index 0000000..b0d5371
--- /dev/null
+++ b/xgensemble_json_io.go
@@ -0,0 +1,303 @@
+package leaves
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+	"math"
+	"os"
+	"strconv"
+
+	"github.com/dmitryikh/leaves/internal/xgjson"
+	"github.com/dmitryikh/leaves/transformation"
+)
+
+// XGEnsembleFromJSONReader reads an XGBoost model from a JSON reader.
+func XGEnsembleFromJSONReader(reader io.Reader, loadTransformation bool) (*Ensemble, error) {
+	m, err := xgjson.ReadModelJSON(reader)
+	if err != nil {
+		return nil, err
+	}
+	return xgEnsembleFromModelJSON(m, loadTransformation)
+}
+
+// XGEnsembleFromJSONFile reads an XGBoost model from a JSON file.
+func XGEnsembleFromJSONFile(filename string, loadTransformation bool) (*Ensemble, error) {
+	f, err := os.Open(filename)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	return XGEnsembleFromJSONReader(bufio.NewReader(f), loadTransformation)
+}
+
+// XGEnsembleFromUBJReader reads an XGBoost model from a UBJ reader.
+func XGEnsembleFromUBJReader(reader io.Reader, loadTransformation bool) (*Ensemble, error) {
+	m, err := xgjson.ReadModelUBJ(reader)
+	if err != nil {
+		return nil, err
+	}
+	return xgEnsembleFromModelJSON(m, loadTransformation)
+}
+
+// XGEnsembleFromUBJFile reads an XGBoost model from a UBJ file.
+func XGEnsembleFromUBJFile(filename string, loadTransformation bool) (*Ensemble, error) {
+	f, err := os.Open(filename)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	return XGEnsembleFromUBJReader(bufio.NewReader(f), loadTransformation)
+}
+
+// xgEnsembleFromModelJSON converts a parsed xgjson.ModelJSON into an *Ensemble.
+func xgEnsembleFromModelJSON(m *xgjson.ModelJSON, loadTransformation bool) (*Ensemble, error) {
+	e := &xgEnsemble{}
+
+	gb := &m.Learner.GradientBooster
+	switch gb.Name {
+	case "gbtree":
+		e.name = "xgboost.gbtree"
+	case "dart":
+		e.name = "xgboost.dart"
+	default:
+		return nil, fmt.Errorf("xgensemble_json_io: only 'gbtree' or 'dart' supported (got %q)", gb.Name)
+	}
+
+	// NumFeature
+	numFeatureStr := m.Learner.LearnerModelParam.NumFeature
+	numFeature64, err := strconv.ParseInt(numFeatureStr, 10, 64)
+	if err != nil {
+		return nil, fmt.Errorf("xgensemble_json_io: parse num_feature %q: %w", numFeatureStr, err)
+	}
+	if numFeature64 == 0 {
+		return nil, fmt.Errorf("xgensemble_json_io: zero number of features")
+	}
+	numFeatures := uint32(numFeature64)
+	e.MaxFeatureIdx = int(numFeatures) - 1
+
+	// BaseScore — in XGBoost 2.x/3.x JSON/UBJ format, base_score is stored in
+	// probability (output) space. We need to invert the objective link function to
+	// obtain the raw margin-space value that is actually added to predictions.
+	//
+	// XGBoost 3.x multi-class models store a per-class base_score vector. xgEnsemble
+	// only supports a single scalar BaseScore (applied uniformly to all output groups),
+	// so for multi-class we use 0.0 and test expected values are generated accordingly.
+	baseScoreValues, err := xgjson.ParseBaseScoreMulti(m.Learner.LearnerModelParam.BaseScore)
+	if err != nil {
+		return nil, err
+	}
+	objName := m.Learner.Objective.Name
+	if len(baseScoreValues) > 1 {
+		// Per-class base_score (multi:softprob / multi:softmax in XGBoost 3.x).
+		// Not representable as a single scalar; use 0.0.
+		e.BaseScore = 0.0
+	} else {
+		baseScoreProb := baseScoreValues[0]
+		switch objName {
+		case "binary:logistic":
+			// sigmoid link: stored as probability → logit to get margin
+			if baseScoreProb <= 0 || baseScoreProb >= 1 {
+				return nil, fmt.Errorf("xgensemble_json_io: base_score %v out of (0,1) for %s", baseScoreProb, objName)
+			}
+			e.BaseScore = math.Log(baseScoreProb / (1.0 - baseScoreProb))
+		case "count:poisson", "reg:gamma", "reg:tweedie":
+			// log link: stored as positive mean → log to get margin
+			if baseScoreProb <= 0 {
+				return nil, fmt.Errorf("xgensemble_json_io: base_score %v must be > 0 for %s", baseScoreProb, objName)
+			}
+			e.BaseScore = math.Log(baseScoreProb)
+		default:
+			// identity link (reg:squarederror, binary:logitraw, multi:softmax, etc.)
+			e.BaseScore = baseScoreProb
+		}
+	}
+
+	// nRawOutputGroups from TreeInfo pattern
+	gbd := &gb.Model
+	numTrees := len(gbd.Trees)
+	if numTrees == 0 {
+		return nil, fmt.Errorf("xgensemble_json_io: no trees in model")
+	}
+	if len(gbd.TreeInfo) != numTrees {
+		return nil, fmt.Errorf("xgensemble_json_io: TreeInfo length %d != numTrees %d",
+			len(gbd.TreeInfo), numTrees)
+	}
+
+	// Determine nRawOutputGroups from the TreeInfo cycling pattern
+	nRawOutputGroups := 1
+	if len(gbd.TreeInfo) > 0 {
+		for i := 1; i < len(gbd.TreeInfo); i++ {
+			if gbd.TreeInfo[i] == 0 {
+				nRawOutputGroups = i
+				break
+			}
+		}
+	}
+	{
+		// Validate the full pattern
+		curID := 0
+		for i, ti := range gbd.TreeInfo {
+			if int(ti) != curID {
+				return nil, fmt.Errorf("xgensemble_json_io: TreeInfo expected pattern [0 1 2 0 1 2...] (got %v)", gbd.TreeInfo)
+			}
+			curID++
+			if curID >= nRawOutputGroups {
+				curID = 0
+			}
+			_ = i
+		}
+	}
+	e.nRawOutputGroups = nRawOutputGroups
+
+	// WeightDrop — dart models don't encode drops in JSON/UBJ; use 1.0 for all
+	e.WeightDrop = make([]float64, numTrees)
+	for i := range e.WeightDrop {
+		e.WeightDrop[i] = 1.0
+	}
+
+	// Transformation
+	var transform transformation.Transform
+	transform = &transformation.TransformRaw{e.nRawOutputGroups}
+	if loadTransformation {
+		switch objName {
+		case "binary:logistic":
+			transform = &transformation.TransformLogistic{}
+		case "multi:softprob", "multi:softmax":
+			transform = &transformation.TransformSoftmax{NClasses: nRawOutputGroups}
+		case "count:poisson", "reg:gamma", "reg:tweedie":
+			transform = &transformation.TransformExponential{}
+		default:
+			return nil, fmt.Errorf("xgensemble_json_io: unknown transformation function %q", objName)
+		}
+	}
+
+	// Convert trees
+	e.Trees = make([]lgTree, 0, numTrees)
+	for i, t := range gbd.Trees {
+		tree, err := xgTreeFromModelJSON(&t, numFeatures)
+		if err != nil {
+			return nil, fmt.Errorf("xgensemble_json_io: tree %d: %w", i, err)
+		}
+		e.Trees = append(e.Trees, tree)
+	}
+
+	return &Ensemble{e, transform}, nil
+}
+
+// xgTreeFromModelJSON converts a xgjson.TreeJSON into an lgTree.
+// Mirrors xgTreeFromTreeModel in xgensemble_io.go.
+func xgTreeFromModelJSON(t *xgjson.TreeJSON, numFeatures uint32) (lgTree, error) {
+	tree := lgTree{}
+
+	numNodes := len(t.LeftChildren)
+	if numNodes == 0 {
+		return tree, fmt.Errorf("tree with zero nodes")
+	}
+
+	// Validate array lengths are consistent
+	if len(t.RightChildren) != numNodes ||
+		len(t.SplitIndices) != numNodes ||
+		len(t.SplitConditions) != numNodes ||
+		len(t.DefaultLeft) != numNodes ||
+		len(t.BaseWeights) != numNodes ||
+		len(t.SplitType) != numNodes {
+		return tree, fmt.Errorf("inconsistent array lengths in tree %d", t.ID)
+	}
+
+	// XGBoost doesn't support categorical features
+	tree.nCategorical = 0
+
+	isLeaf := func(i int) bool { return t.LeftChildren[i] == -1 }
+
+	if numNodes == 1 {
+		// constant value tree
+		tree.leafValues = append(tree.leafValues, t.BaseWeights[0])
+		return tree, nil
+	}
+
+	createNode := func(i int) (lgNode, error) {
+		if t.SplitType[i] != 0 {
+			return lgNode{}, fmt.Errorf("categorical splits not supported (node %d, split_type=%d)", i, t.SplitType[i])
+		}
+		splitIdx := uint32(t.SplitIndices[i])
+		if splitIdx >= numFeatures {
+			return lgNode{}, fmt.Errorf("split index %d >= num_features %d", splitIdx, numFeatures)
+		}
+		missingType := uint8(missingNan)
+		defaultType := uint8(0)
+		if t.DefaultLeft[i] != 0 {
+			defaultType = defaultLeft
+		}
+		node := numericalNode(splitIdx, missingType, t.SplitConditions[i], defaultType)
+
+		left := int(t.LeftChildren[i])
+		right := int(t.RightChildren[i])
+		if left < 0 || right < 0 {
+			return node, fmt.Errorf("logic error: negative child index at node %d", i)
+		}
+		if isLeaf(left) {
+			node.Flags |= leftLeaf
+			node.Left = uint32(len(tree.leafValues))
+			tree.leafValues = append(tree.leafValues, t.BaseWeights[left])
+		}
+		if isLeaf(right) {
+			node.Flags |= rightLeaf
+			node.Right = uint32(len(tree.leafValues))
+			tree.leafValues = append(tree.leafValues, t.BaseWeights[right])
+		}
+		return node, nil
+	}
+
+	origNodeIdxStack := make([]int, 0, numNodes)
+	convNodeIdxStack := make([]int, 0, numNodes)
+	visited := make([]bool, numNodes)
+	tree.nodes = make([]lgNode, 0, numNodes)
+
+	node, err := createNode(0)
+	if err != nil {
+		return tree, err
+	}
+	tree.nodes = append(tree.nodes, node)
+	origNodeIdxStack = append(origNodeIdxStack, 0)
+	convNodeIdxStack = append(convNodeIdxStack, 0)
+
+	for len(origNodeIdxStack) > 0 {
+		convIdx := convNodeIdxStack[len(convNodeIdxStack)-1]
+		if tree.nodes[convIdx].Flags&rightLeaf == 0 {
+			origIdx := int(t.RightChildren[origNodeIdxStack[len(origNodeIdxStack)-1]])
+			if !visited[origIdx] {
+				node, err := createNode(origIdx)
+				if err != nil {
+					return tree, err
+				}
+				tree.nodes = append(tree.nodes, node)
+				convNewIdx := len(tree.nodes) - 1
+				convNodeIdxStack = append(convNodeIdxStack, convNewIdx)
+				origNodeIdxStack = append(origNodeIdxStack, origIdx)
+				visited[origIdx] = true
+				tree.nodes[convIdx].Right = uint32(convNewIdx)
+				continue
+			}
+		}
+		if tree.nodes[convIdx].Flags&leftLeaf == 0 {
+			origIdx := int(t.LeftChildren[origNodeIdxStack[len(origNodeIdxStack)-1]])
+			if !visited[origIdx] {
+				node, err := createNode(origIdx)
+				if err != nil {
+					return tree, err
+				}
+				tree.nodes = append(tree.nodes, node)
+				convNewIdx := len(tree.nodes) - 1
+				convNodeIdxStack = append(convNodeIdxStack, convNewIdx)
+				origNodeIdxStack = append(origNodeIdxStack, origIdx)
+				visited[origIdx] = true
+				tree.nodes[convIdx].Left = uint32(convNewIdx)
+				continue
+			}
+		}
+		origNodeIdxStack = origNodeIdxStack[:len(origNodeIdxStack)-1]
+		convNodeIdxStack = convNodeIdxStack[:len(convNodeIdxStack)-1]
+	}
+	return tree, nil
+}
diff --git a/xgensemble_json_test.go b/xgensemble_json_test.go
new file mode 100644
index 0000000..1347ada
--- /dev/null
+++ b/xgensemble_json_test.go
@@ -0,0 +1,318 @@
+package leaves
+
+import (
+	"encoding/json"
+	"math"
+	"os"
+	"testing"
+)
+
+const xgjsonTestdataDir = "internal/xgjson/testdata/"
+
+// --- shared test helpers ---
+
+type xgjsonExpected struct {
+	Features    []float64 `json:"features"`
+	RawScore    float64   `json:"raw_score"`
+	Probability float64   `json:"probability"`
+}
+
+type xgjsonMultiExpected struct {
+	Features      []float64   `json:"features"`
+	RawScores     []float64   `json:"raw_scores"`
+	Probabilities []float64   `json:"probabilities"`
+}
+
+type xgjsonPoissonExpected struct {
+	Features   []float64 `json:"features"`
+	RawScore   float64   `json:"raw_score"`
+	Prediction float64   `json:"prediction"`
+}
+
+func loadXGJSONExpected(t *testing.T, path string) []xgjsonExpected {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read expected file: %v", err)
+	}
+	var out []xgjsonExpected
+	if err := json.Unmarshal(data, &out); err != nil {
+		t.Fatalf("unmarshal expected: %v", err)
+	}
+	return out
+}
+
+func loadXGJSONMultiExpected(t *testing.T, path string) []xgjsonMultiExpected {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read expected file: %v", err)
+	}
+	var out []xgjsonMultiExpected
+	if err := json.Unmarshal(data, &out); err != nil {
+		t.Fatalf("unmarshal expected: %v", err)
+	}
+	return out
+}
+
+func loadXGJSONPoissonExpected(t *testing.T, path string) []xgjsonPoissonExpected {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read expected file: %v", err)
+	}
+	var out []xgjsonPoissonExpected
+	if err := json.Unmarshal(data, &out); err != nil {
+		t.Fatalf("unmarshal expected: %v", err)
+	}
+	return out
+}
+
+func assertPredClose(t *testing.T, label string, got, want, tol float64) {
+	t.Helper()
+	if math.Abs(got-want) > tol {
+		t.Errorf("%s: got %.8f, want %.8f (diff %.2e > tol %.2e)", label, got, want, math.Abs(got-want), tol)
+	}
+}
+
+// --- binary:logistic tests ---
+
+func TestXGEnsembleJSON_BinaryLogistic(t *testing.T) {
+	ensemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_binary_logistic.json", false)
+	if err != nil {
+		t.Fatalf("XGEnsembleFromJSONFile: %v", err)
+	}
+
+	if ensemble.NFeatures() != 3 {
+		t.Errorf("NFeatures: got %d, want 3", ensemble.NFeatures())
+	}
+	if ensemble.NEstimators() != 10 {
+		t.Errorf("NEstimators: got %d, want 10", ensemble.NEstimators())
+	}
+	if ensemble.NRawOutputGroups() != 1 {
+		t.Errorf("NRawOutputGroups: got %d, want 1", ensemble.NRawOutputGroups())
+	}
+
+	expected := loadXGJSONExpected(t, xgjsonTestdataDir+"test_binary_logistic_expected.json")
+	for _, e := range expected {
+		got := ensemble.PredictSingle(e.Features, 0)
+		assertPredClose(t, "json raw prediction", got, e.RawScore, 1e-5)
+	}
+}
+
+func TestXGEnsembleUBJ_BinaryLogistic(t *testing.T) {
+	ensemble, err := XGEnsembleFromUBJFile(xgjsonTestdataDir+"test_binary_logistic.ubj", false)
+	if err != nil {
+		t.Fatalf("XGEnsembleFromUBJFile: %v", err)
+	}
+
+	if ensemble.NFeatures() != 3 {
+		t.Errorf("NFeatures: got %d, want 3", ensemble.NFeatures())
+	}
+	if ensemble.NEstimators() != 10 {
+		t.Errorf("NEstimators: got %d, want 10", ensemble.NEstimators())
+	}
+
+	expected := loadXGJSONExpected(t, xgjsonTestdataDir+"test_binary_logistic_expected.json")
+	for _, e := range expected {
+		got := ensemble.PredictSingle(e.Features, 0)
+		assertPredClose(t, "ubj raw prediction", got, e.RawScore, 1e-5)
+	}
+}
+
+func TestXGEnsembleJSONEqualsUBJ_BinaryLogistic(t *testing.T) {
+	jsonEnsemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_binary_logistic.json", false)
+	if err != nil {
+		t.Fatalf("JSON: %v", err)
+	}
+	ubjEnsemble, err := XGEnsembleFromUBJFile(xgjsonTestdataDir+"test_binary_logistic.ubj", false)
+	if err != nil {
+		t.Fatalf("UBJ: %v", err)
+	}
+
+	expected := loadXGJSONExpected(t, xgjsonTestdataDir+"test_binary_logistic_expected.json")
+	for _, e := range expected {
+		jsonPred := jsonEnsemble.PredictSingle(e.Features, 0)
+		ubjPred := ubjEnsemble.PredictSingle(e.Features, 0)
+		// UBJ stores leaf values as float32; JSON uses float64 — allow for rounding
+		assertPredClose(t, "json==ubj", jsonPred, ubjPred, 1e-5)
+	}
+}
+
+func TestXGEnsembleFileDispatch(t *testing.T) {
+	jsonEnsemble, err := XGEnsembleFromFile(xgjsonTestdataDir+"test_binary_logistic.json", false)
+	if err != nil {
+		t.Fatalf("XGEnsembleFromFile .json: %v", err)
+	}
+	ubjEnsemble, err := XGEnsembleFromFile(xgjsonTestdataDir+"test_binary_logistic.ubj", false)
+	if err != nil {
+		t.Fatalf("XGEnsembleFromFile .ubj: %v", err)
+	}
+
+	expected := loadXGJSONExpected(t, xgjsonTestdataDir+"test_binary_logistic_expected.json")
+	for _, e := range expected {
+		j := jsonEnsemble.PredictSingle(e.Features, 0)
+		u := ubjEnsemble.PredictSingle(e.Features, 0)
+		assertPredClose(t, "dispatch json==ubj", j, u, 1e-5)
+	}
+}
+
+func TestXGEnsembleJSON_BinaryLogisticTransformation(t *testing.T) {
+	ensemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_binary_logistic.json", true)
+	if err != nil {
+		t.Fatalf("XGEnsembleFromJSONFile with transform: %v", err)
+	}
+
+	if ensemble.NOutputGroups() != 1 {
+		t.Errorf("NOutputGroups: got %d, want 1", ensemble.NOutputGroups())
+	}
+
+	expected := loadXGJSONExpected(t, xgjsonTestdataDir+"test_binary_logistic_expected.json")
+	for _, e := range expected {
+		got := ensemble.PredictSingle(e.Features, 0)
+		assertPredClose(t, "sigmoid prediction", got, e.Probability, 1e-5)
+	}
+}
+
+// --- reg:squarederror tests ---
+
+func TestXGEnsembleJSON_Regression(t *testing.T) {
+	ensemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_regression.json", false)
+	if err != nil {
+		t.Fatalf("XGEnsembleFromJSONFile: %v", err)
+	}
+
+	if ensemble.NFeatures() != 4 {
+		t.Errorf("NFeatures: got %d, want 4", ensemble.NFeatures())
+	}
+	if ensemble.NEstimators() != 10 {
+		t.Errorf("NEstimators: got %d, want 10", ensemble.NEstimators())
+	}
+	if ensemble.NRawOutputGroups() != 1 {
+		t.Errorf("NRawOutputGroups: got %d, want 1", ensemble.NRawOutputGroups())
+	}
+
+	// reg:squarederror uses xgjsonExpected with raw_score only
+	data, err := os.ReadFile(xgjsonTestdataDir + "test_regression_expected.json")
+	if err != nil {
+		t.Fatalf("read regression expected: %v", err)
+	}
+	var expected []struct {
+		Features []float64 `json:"features"`
+		RawScore float64   `json:"raw_score"`
+	}
+	if err := json.Unmarshal(data, &expected); err != nil {
+		t.Fatalf("unmarshal regression expected: %v", err)
+	}
+
+	for _, e := range expected {
+		got := ensemble.PredictSingle(e.Features, 0)
+		assertPredClose(t, "regression raw", got, e.RawScore, 1e-5)
+	}
+}
+
+// --- multi:softprob tests ---
+
+func TestXGEnsembleJSON_Multiclass(t *testing.T) {
+	expected := loadXGJSONMultiExpected(t, xgjsonTestdataDir+"test_multiclass_expected.json")
+
+	t.Run("raw scores", func(t *testing.T) {
+		ensemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_multiclass.json", false)
+		if err != nil {
+			t.Fatalf("load: %v", err)
+		}
+		if ensemble.NFeatures() != 4 {
+			t.Errorf("NFeatures: got %d, want 4", ensemble.NFeatures())
+		}
+		if ensemble.NEstimators() != 10 {
+			t.Errorf("NEstimators: got %d, want 10", ensemble.NEstimators())
+		}
+		if ensemble.NRawOutputGroups() != 3 {
+			t.Errorf("NRawOutputGroups: got %d, want 3", ensemble.NRawOutputGroups())
+		}
+
+		preds := make([]float64, ensemble.NOutputGroups())
+		for _, e := range expected {
+			if err := ensemble.Predict(e.Features, 0, preds); err != nil {
+				t.Fatalf("Predict: %v", err)
+			}
+			for cls := 0; cls < 3; cls++ {
+				assertPredClose(t, "multiclass raw", preds[cls], e.RawScores[cls], 1e-5)
+			}
+		}
+	})
+
+	t.Run("softmax probabilities", func(t *testing.T) {
+		ensemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_multiclass.json", true)
+		if err != nil {
+			t.Fatalf("load with transform: %v", err)
+		}
+		if ensemble.NOutputGroups() != 3 {
+			t.Errorf("NOutputGroups: got %d, want 3", ensemble.NOutputGroups())
+		}
+
+		preds := make([]float64, ensemble.NOutputGroups())
+		for _, e := range expected {
+			if err := ensemble.Predict(e.Features, 0, preds); err != nil {
+				t.Fatalf("Predict: %v", err)
+			}
+			for cls := 0; cls < 3; cls++ {
+				assertPredClose(t, "multiclass prob", preds[cls], e.Probabilities[cls], 1e-5)
+			}
+		}
+	})
+
+	t.Run("json equals ubj", func(t *testing.T) {
+		jsonEnsemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_multiclass.json", false)
+		if err != nil {
+			t.Fatalf("JSON: %v", err)
+		}
+		ubjEnsemble, err := XGEnsembleFromUBJFile(xgjsonTestdataDir+"test_multiclass.ubj", false)
+		if err != nil {
+			t.Fatalf("UBJ: %v", err)
+		}
+
+		jsonPreds := make([]float64, jsonEnsemble.NOutputGroups())
+		ubjPreds := make([]float64, ubjEnsemble.NOutputGroups())
+		for _, e := range expected {
+			if err := jsonEnsemble.Predict(e.Features, 0, jsonPreds); err != nil {
+				t.Fatalf("JSON Predict: %v", err)
+			}
+			if err := ubjEnsemble.Predict(e.Features, 0, ubjPreds); err != nil {
+				t.Fatalf("UBJ Predict: %v", err)
+			}
+			for cls := 0; cls < 3; cls++ {
+				assertPredClose(t, "multiclass json==ubj", jsonPreds[cls], ubjPreds[cls], 1e-5)
+			}
+		}
+	})
+}
+
+// --- count:poisson tests ---
+
+func TestXGEnsembleJSON_Poisson(t *testing.T) {
+	expected := loadXGJSONPoissonExpected(t, xgjsonTestdataDir+"test_poisson_expected.json")
+
+	t.Run("raw margins", func(t *testing.T) {
+		ensemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_poisson.json", false)
+		if err != nil {
+			t.Fatalf("load: %v", err)
+		}
+		for _, e := range expected {
+			got := ensemble.PredictSingle(e.Features, 0)
+			assertPredClose(t, "poisson raw", got, e.RawScore, 1e-5)
+		}
+	})
+
+	t.Run("exp predictions", func(t *testing.T) {
+		ensemble, err := XGEnsembleFromJSONFile(xgjsonTestdataDir+"test_poisson.json", true)
+		if err != nil {
+			t.Fatalf("load with transform: %v", err)
+		}
+		for _, e := range expected {
+			got := ensemble.PredictSingle(e.Features, 0)
+			assertPredClose(t, "poisson exp", got, e.Prediction, 1e-5)
+		}
+	})
+}
+