Summary

A proof of concept for in-process fine-tuning, wiring llama.cpp's ggml-opt training path into
the JNI layer. It mirrors upstream examples/training/finetune.cpp: load a model, tokenize a text
corpus into a ggml-opt dataset, run llama_opt_init + llama_opt_epoch for N epochs, and write the
fine-tuned GGUF via llama_model_save_to_file.

New API:

LlamaTrainer.finetune(Path model, String trainingText, Path output, int epochs, float learningRate);

Why this is a small change

The open question for "can java-llama.cpp train like llama.cpp?" was whether the
ggml-opt / llama_opt machinery even links into our static libjllama. It does — verified with
nm on the built library: llama_opt_init, llama_opt_epoch, ggml_opt_fit,
ggml_opt_dataset_init, common_opt_dataset_init, common_opt_lr_pars, and
llama_model_save_to_file are all defined (T) symbols, and CMake already links llama-common. So
this is pure JNI + C++ wiring with no build-system change.

What's verified

• train_engine.cpp compiles and links into libjllama (b9842).
• The finetuneNative JNI symbol is exported; the library loads cleanly (NativeLibraryLoadSmokeTest).
• LlamaTrainer compiles through the strict Error Prone / NullAway pipeline.
• The actual training run is exercised by LlamaTrainerIntegrationTest, which self-skips unless
  -Dnet.ladenthin.llama.train.model=/path/to/small.gguf is set (full-model fine-tuning is
  compute/memory-heavy and should not run in a default build).

Design

• train_engine.{h,cpp} — a self-contained native finetune(), independent of the inference
  server_context; it loads its own model + context and forces the two settings upstream training
  requires (no mmap → writable weights; f32 KV cache → OUT_PROD has no f16 support). It
  intentionally does not call llama_backend_free(), since other live contexts in the JVM may
  still depend on the initialized backend.
• LlamaTrainer — a deliberately minimal Java surface so the native path can be exercised before a
  richer API is designed.

Scope / next steps

This is a POC, not a finished feature — upstream training support is itself experimental (full or
selective fine-tune, small models). A follow-up FineTuner API could add: dataset/file input and
batching, optimizer and LoRA-target selection, a learning-rate schedule, validation split, and
progress callbacks. Opening this to confirm the approach and the in-process feasibility.