|
| 1 | +"""Compute per-turn mean NLL of tool output tokens for SWE-bench trajectories. |
| 2 | +
|
| 3 | +For each assistant turn k > 0, the model's mean negative log-likelihood over the |
| 4 | +tool/environment output tokens that precede turn k is stored. This quantifies how |
| 5 | +"surprised" the model was by the test runner or compiler response. |
| 6 | +
|
| 7 | +Output: <output> (a .pt file) |
| 8 | + = {sample_id: {turn_k (int): mean_nll (float)}} |
| 9 | +
|
| 10 | +Turn 0 is excluded because it has no preceding tool output (only the initial task). |
| 11 | +
|
| 12 | +Usage (array job, 8 shards): |
| 13 | + sbatch --array=0-7 slurm/build_tool_nll.sh \\ |
| 14 | + --model-config configs/models/laguna_xs2.yaml \\ |
| 15 | + --generation-config configs/generation_laguna_xs2.yaml \\ |
| 16 | + --traj-dir generations/swebench/laguna_xs2_full \\ |
| 17 | + --output cache/swebench/laguna_xs2_full/tool_nll_index.pt |
| 18 | +""" |
| 19 | +from __future__ import annotations |
| 20 | + |
| 21 | +import argparse |
| 22 | +from pathlib import Path |
| 23 | + |
| 24 | +import torch |
| 25 | + |
| 26 | +from src.configs import GenerationConfig, ModelConfig, load_config |
| 27 | +from src.tasks.swe_bench_extract import load_trajectories |
| 28 | + |
| 29 | + |
| 30 | +def compute_trajectory_nll( |
| 31 | + token_ids: list[int], |
| 32 | + segments: list[dict], |
| 33 | + step_segment_indices: list[int], |
| 34 | + logits: torch.Tensor, |
| 35 | +) -> dict[int, float]: |
| 36 | + """Per-turn mean NLL of tool output tokens for turns k > 0. |
| 37 | +
|
| 38 | + logits[i] is the model's prediction for token i+1, so the NLL of a tool |
| 39 | + output spanning token positions [start, end) is: |
| 40 | + mean(-log softmax(logits[start-1 : end-1])[:, token_ids[start:end]]) |
| 41 | +
|
| 42 | + Args: |
| 43 | + token_ids: full token sequence (length = seq_len) |
| 44 | + segments: tokenization.segments list (role, start_token, end_token) |
| 45 | + step_segment_indices: indices into *segments* for each assistant turn |
| 46 | + logits: [seq_len, vocab_size] — output of one full forward pass |
| 47 | +
|
| 48 | + Returns: |
| 49 | + {turn_k: mean_nll} for turns that have a preceding user (tool output) segment |
| 50 | + """ |
| 51 | + turn_nll: dict[int, float] = {} |
| 52 | + for k, asst_seg_idx in enumerate(step_segment_indices): |
| 53 | + if k == 0 or asst_seg_idx == 0: |
| 54 | + continue # turn 0 is preceded by the initial task prompt, not a tool response |
| 55 | + tool_seg = segments[asst_seg_idx - 1] |
| 56 | + if tool_seg.get("role") != "user": |
| 57 | + continue |
| 58 | + start = tool_seg["start_token"] |
| 59 | + end = tool_seg["end_token"] |
| 60 | + n_tokens = end - start |
| 61 | + if n_tokens <= 0 or start == 0: |
| 62 | + continue |
| 63 | + # logits[start-1] predicts token[start], …, logits[end-2] predicts token[end-1] |
| 64 | + tool_logits = logits[start - 1 : end - 1] # [n_tokens, vocab_size] |
| 65 | + if tool_logits.shape[0] != n_tokens: |
| 66 | + continue |
| 67 | + actual = torch.tensor(token_ids[start:end], dtype=torch.long, device=logits.device) |
| 68 | + log_probs = torch.log_softmax(tool_logits.float(), dim=-1) |
| 69 | + token_nll = -log_probs[torch.arange(n_tokens, device=logits.device), actual] |
| 70 | + turn_nll[k] = float(token_nll.mean().item()) |
| 71 | + return turn_nll |
| 72 | + |
| 73 | + |
| 74 | +def _load_model_adapter(adapter_name: str): |
| 75 | + if adapter_name == "laguna": |
| 76 | + from src.models.laguna import LagunaAdapter |
| 77 | + return LagunaAdapter() |
| 78 | + if adapter_name == "qwen": |
| 79 | + from src.models.qwen import QwenAdapter |
| 80 | + return QwenAdapter() |
| 81 | + if adapter_name == "qwen35": |
| 82 | + from src.models.qwen35 import Qwen35Adapter |
| 83 | + return Qwen35Adapter() |
| 84 | + if adapter_name == "cwm": |
| 85 | + from src.models.cwm import CwmAdapter |
| 86 | + return CwmAdapter() |
| 87 | + raise ValueError(f"Unknown model adapter: {adapter_name!r}") |
| 88 | + |
| 89 | + |
| 90 | +def main() -> None: |
| 91 | + parser = argparse.ArgumentParser(description=__doc__) |
| 92 | + parser.add_argument("--model-config", required=True) |
| 93 | + parser.add_argument("--generation-config", required=True) |
| 94 | + parser.add_argument("--traj-dir", required=True, help="Directory of trajectory JSON files") |
| 95 | + parser.add_argument("--output", required=True, help="Output .pt path") |
| 96 | + parser.add_argument("--shard-rank", type=int, default=0) |
| 97 | + parser.add_argument("--num-shards", type=int, default=1) |
| 98 | + args = parser.parse_args() |
| 99 | + |
| 100 | + model_config: ModelConfig = load_config(args.model_config, ModelConfig) |
| 101 | + gen_config: GenerationConfig = load_config(args.generation_config, GenerationConfig) |
| 102 | + |
| 103 | + out_path = Path(args.output) |
| 104 | + # If sharded, accumulate into a shard-specific temp file; merge manually afterwards. |
| 105 | + # For simplicity we write the whole shard's results to the output path (caller merges). |
| 106 | + if args.num_shards > 1: |
| 107 | + out_path = out_path.with_suffix(f".shard{args.shard_rank}.pt") |
| 108 | + |
| 109 | + print(f"Loading trajectories from {args.traj_dir}...") |
| 110 | + all_trajs = load_trajectories(args.traj_dir) |
| 111 | + trajs = all_trajs[args.shard_rank::args.num_shards] |
| 112 | + print(f" {len(trajs)} trajectories (shard {args.shard_rank}/{args.num_shards})") |
| 113 | + |
| 114 | + model_adapter = _load_model_adapter(model_config.adapter) |
| 115 | + model_adapter.load_for_extraction(model_config, gen_config) |
| 116 | + device = next(model_adapter._model.parameters()).device |
| 117 | + |
| 118 | + index: dict[str, dict[int, float]] = {} |
| 119 | + |
| 120 | + for i, traj in enumerate(trajs): |
| 121 | + print(f" [{i + 1}/{len(trajs)}] {traj.sample_id} ({len(traj.token_ids)} tokens)...", flush=True) |
| 122 | + input_ids = torch.tensor([traj.token_ids], dtype=torch.long, device=device) |
| 123 | + try: |
| 124 | + with torch.no_grad(): |
| 125 | + out = model_adapter._model(input_ids=input_ids, output_hidden_states=False) |
| 126 | + logits = out.logits[0].cpu() # [seq_len, vocab_size] — move to CPU immediately |
| 127 | + del out |
| 128 | + torch.cuda.empty_cache() |
| 129 | + except torch.cuda.OutOfMemoryError: |
| 130 | + torch.cuda.empty_cache() |
| 131 | + print(f" OOM — skipping {traj.sample_id}") |
| 132 | + continue |
| 133 | + |
| 134 | + turn_nll = compute_trajectory_nll( |
| 135 | + traj.token_ids, traj.segments, traj.step_segment_indices, logits |
| 136 | + ) |
| 137 | + del logits |
| 138 | + index[traj.sample_id] = turn_nll |
| 139 | + print(f" {len(turn_nll)} turns with tool NLL (out of {len(traj.step_segment_indices)} total turns)") |
| 140 | + |
| 141 | + out_path.parent.mkdir(parents=True, exist_ok=True) |
| 142 | + torch.save(index, out_path) |
| 143 | + print(f"Saved tool NLL index ({len(index)} samples) → {out_path}") |
| 144 | + |
| 145 | + |
| 146 | +if __name__ == "__main__": |
| 147 | + main() |
0 commit comments