#!/usr/bin/env python3 # -*- coding: utf-8 -*- """合并分块 analysis 的 neuron_analysis: 逐 (module,layer) 把各 chunk 的 per-channel 高激活 entry list 拼接 (sentence_idx 按累计偏移修正以支持 ctx), sentence 拼接。结果等价于在全部 样本上一次性 analysis (na-count = 各 chunk 之和, top-K-by-activation 不变)。 用法: merge_chunks.py ... """ import json, sys, glob from pathlib import Path def _load(p): try: import orjson return orjson.loads(Path(p).read_bytes()) except Exception: return json.loads(Path(p).read_text(encoding="utf-8")) def main(out_dir, chunk_dirs): out_dir = Path(out_dir); out_dir.mkdir(parents=True, exist_ok=True) chunk_dirs = [Path(c) for c in chunk_dirs] bases = sorted(p.name.replace("_neuron_analysis.json", "") for p in chunk_dirs[0].glob("*_neuron_analysis.json")) print(f"merging {len(chunk_dirs)} chunks, {len(bases)} module-layer files", flush=True) for bi, base in enumerate(bases): merged = {}; merged_sents = []; offset = 0 for cd in chunk_dirs: naf = cd / f"{base}_neuron_analysis.json" sf = cd / f"{base}_sentence.json" if not naf.exists(): continue ntt = _load(naf).get("neuron_top_tokens", {}) sents = _load(sf).get("sentences", []) if sf.exists() else [] for ch, entries in ntt.items(): if offset: for e in entries: if isinstance(e, dict) and e.get("sentence_idx") is not None: e["sentence_idx"] = e["sentence_idx"] + offset merged.setdefault(ch, []).extend(entries) merged_sents.extend(sents); offset += len(sents) (out_dir / f"{base}_neuron_analysis.json").write_text( json.dumps({"neuron_top_tokens": merged}, ensure_ascii=False, separators=(",", ":")), encoding="utf-8") (out_dir / f"{base}_sentence.json").write_text( json.dumps({"sentences": merged_sents}, ensure_ascii=False, separators=(",", ":")), encoding="utf-8") if (bi + 1) % 20 == 0: print(f" {bi+1}/{len(bases)} done (last {base}: {offset} sents)", flush=True) print(f"merged -> {out_dir}", flush=True) if __name__ == "__main__": main(sys.argv[1], sys.argv[2:])