"""Baseline refusal behaviour per candidate category and language, and the model's own refusal / compliance onsets (needed to label the refusal EAP task). For every pool prompt with a text in a language we generate 48 tokens (shared chat format), run AuditKIT's RefusalDetector, and record the first generated token. Outputs data/b2/baseline_.jsonl and data/b2/baseline_summary.json: refusal rate harmful vs benign per category x language, and the top-5 first tokens of refusals and of compliances per language.""" from __future__ import annotations import json import sys from collections import Counter, defaultdict from pathlib import Path sys.path.insert(0, str(Path(__file__).resolve().parent.parent)) import torch from aya_circuits.config import load_config from aya_circuits.model import chat_wrap, load_hf from neb.judge import refused as judge_refused HERE = Path(__file__).resolve().parent / "b2" @torch.no_grad() def gen_with_first(model, tok, texts, max_new, bs): tok.padding_side = "left" outs, firsts = [], [] for i in range(0, len(texts), bs): enc = tok( texts[i : i + bs], return_tensors="pt", padding=True, add_special_tokens=False ).to(model.device) g = model.generate( **enc, max_new_tokens=max_new, do_sample=False, pad_token_id=tok.pad_token_id ) new = g[:, enc["input_ids"].shape[1] :] outs += tok.batch_decode(new, skip_special_tokens=True) firsts += new[:, 0].tolist() return outs, firsts def main() -> None: cfg = load_config() bs, max_new = int(cfg.get("batch_size", 16)), int(cfg["audit"]["max_new_tokens"]) model, tok = load_hf(cfg) summary, onsets = {}, defaultdict(lambda: {"refusal": Counter(), "comply": Counter()}) for pool in sorted(HERE.glob("pool_*.jsonl")): cat = pool.stem[5:] rows = [json.loads(line) for line in pool.open()] rates = {} for lang in cfg["languages"]: idx = [i for i, r in enumerate(rows) if lang in r] if not idx: continue texts = [chat_wrap(tok, rows[i][lang], cfg["chat_format"]) for i in idx] outs, firsts = gen_with_first(model, tok, texts, max_new, bs) for i, o, f in zip(idx, outs, firsts, strict=False): refused = judge_refused(rows[i][lang], o, lang) rows[i].setdefault("gen", {})[lang] = { "first_token": f, "refused": refused, "preview": o[:160], } onsets[lang]["refusal" if refused else "comply"][tok.decode([f])] += 1 for split in ("harmful", "benign"): sel = [rows[i]["gen"][lang]["refused"] for i in idx if rows[i]["split"] == split] if sel: rates[f"{lang}_{split}"] = { "refusal_rate": round(sum(sel) / len(sel), 3), "n": len(sel), } print(cat, lang, {k: v for k, v in rates.items() if k.startswith(lang)}, flush=True) # raw rows (with harmful-reply previews) stay local; the repo copy is sanitized (brief §8) from data.sanitize import PRIVATE, strip_baseline_row PRIVATE.mkdir(parents=True, exist_ok=True) (PRIVATE / f"baseline_{cat}.jsonl").write_text( "".join(json.dumps(r, ensure_ascii=False) + "\n" for r in rows) ) with (HERE / f"baseline_{cat}.jsonl").open("w") as f: for r in rows: f.write(json.dumps(strip_baseline_row(r), ensure_ascii=False) + "\n") summary[cat] = rates summary["onsets"] = { lang: {k: c.most_common(6) for k, c in v.items()} for lang, v in onsets.items() } (HERE / "baseline_summary.json").write_text(json.dumps(summary, indent=2, ensure_ascii=False)) print(json.dumps(summary["onsets"], ensure_ascii=False)) def rejudge() -> None: """Recompute `refused`, onsets and the summary from stored previews (no model). Harmful-row previews live only in the local raw copy (data/sanitize.py, brief §8); the repo file is rewritten sanitized.""" from transformers import AutoTokenizer from data.sanitize import PRIVATE, strip_baseline_row cfg = load_config() tok = AutoTokenizer.from_pretrained(cfg["model"]) summary, onsets = {}, defaultdict(lambda: {"refusal": Counter(), "comply": Counter()}) for pool in sorted(HERE.glob("baseline_*.jsonl")): cat = pool.stem[9:] raw = PRIVATE / pool.name if (PRIVATE / pool.name).exists() else pool rows = [json.loads(line) for line in raw.open()] rates = {} for lang in cfg["languages"]: for r in rows: g = r.get("gen", {}).get(lang) if g: g["refused"] = judge_refused(r[lang], g["preview"], lang) onsets[lang]["refusal" if g["refused"] else "comply"][ tok.decode([g["first_token"]]) ] += 1 for split in ("harmful", "benign"): sel = [ r["gen"][lang]["refused"] for r in rows if r["split"] == split and lang in r.get("gen", {}) ] if sel: rates[f"{lang}_{split}"] = { "refusal_rate": round(sum(sel) / len(sel), 3), "n": len(sel), } if raw != pool: raw.write_text("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)) with pool.open("w") as f: for r in rows: f.write(json.dumps(strip_baseline_row(r), ensure_ascii=False) + "\n") summary[cat] = rates summary["onsets"] = { lang: {k: c.most_common(6) for k, c in v.items()} for lang, v in onsets.items() } summary["note"] = "judged from 160-char previews with audit/judge.py" (HERE / "baseline_summary.json").write_text(json.dumps(summary, indent=2, ensure_ascii=False)) print( json.dumps( {k: v for k, v in summary.items() if k != "onsets"}, ensure_ascii=False, indent=1 ) ) print(json.dumps(summary["onsets"], ensure_ascii=False)) if __name__ == "__main__": ( rejudge() if sys.argv.pop(sys.argv.index("--rejudge")) else None ) if "--rejudge" in sys.argv else main()