Download b2_build.py from MdSadiqMd/circuit-steering-data: direct link, hf CLI and curl.
- Browser
- Download file 4.77 kB
-
https://huggingface.co/datasets/MdSadiqMd/circuit-steering-data/resolve/main/b2_build.py
- Command line
-
hf download hf://datasets/MdSadiqMd/circuit-steering-data/b2_build.py
-
curl -L -o b2_build.py https://huggingface.co/datasets/MdSadiqMd/circuit-steering-data/resolve/main/b2_build.py
4.77 kB
| """B2 candidate pools: harmful-request categories with matched benign near-miss prompts (English). | |
| Sources (prompts only; no generated harmful content is ever stored, brief §8): | |
| HarmBench standard behaviours (centerforaisafety/HarmBench, CSV) -> semantic categories | |
| MultiJail (DAMO-NLP-SG, human-translated en/ar/sw + 7 more) -> tag categories, gives | |
| human ar and sw for | |
| free | |
| XSTest safe prompts (paul-rottger/xstest, CSV) -> near-miss benign | |
| Three candidate categories are built; the one used for B2 is chosen after measuring Tiny Aya's | |
| baseline refusal per language (data/filter_model.py) and recorded in docs/preregistration.md. | |
| Translations into hi/fr (and ar/sw where MultiJail has none) are produced by data/translate.py. | |
| Output: data/b2/pool_<category>.jsonl with fields id, source, split(harmful|benign), en[, ar, sw]. | |
| """ | |
| from __future__ import annotations | |
| import ast | |
| import csv | |
| import json | |
| from pathlib import Path | |
| RAW = Path(__file__).resolve().parent.parent / ".cache/raw" | |
| OUT = Path(__file__).resolve().parent / "b2" | |
| # category -> (HarmBench SemanticCategory set, MultiJail tag set, XSTest focus/note keywords for | |
| # near-miss) | |
| CATEGORIES = { | |
| "cyber": ( | |
| {"cybercrime_intrusion"}, | |
| set(), | |
| {"hack", "steal", "exploit", "attack", "crack", "kill", "terminate", "execute"}, | |
| ), | |
| "weapons_violence": ( | |
| {"illegal", "harmful"}, | |
| {"Weapons", "Violence & incitement", "Terrorism & organized crime"}, | |
| {"violence", "weapon", "gun", "bomb", "explode", "shoot", "fire", "blow", "knife"}, | |
| ), | |
| "theft_fraud": ( | |
| set(), | |
| {"Theft", "Fraud & deception"}, | |
| {"steal", "theft", "money", "scam", "fraud", "bank", "pick", "lock"}, | |
| ), | |
| } | |
| def main() -> None: | |
| OUT.mkdir(parents=True, exist_ok=True) | |
| hb = [ | |
| r | |
| for r in csv.DictReader((RAW / "harmbench/harmbench_behaviors_text_all.csv").open()) | |
| if r["FunctionalCategory"] == "standard" | |
| ] | |
| mj = list(csv.DictReader((RAW / "MultiJail/MultiJail.csv").open(encoding="utf-8-sig"))) | |
| xs = list(csv.DictReader((RAW / "xstest/xstest_prompts.csv").open())) | |
| safe = [r for r in xs if r["label"] == "safe"] | |
| stats = {} | |
| for cat, (hb_cats, mj_tags, kws) in CATEGORIES.items(): | |
| rows = [ | |
| { | |
| "id": f"hb:{r['BehaviorID']}", | |
| "source": "harmbench", | |
| "split": "harmful", | |
| "en": r["Behavior"], | |
| } | |
| for r in hb | |
| if r["SemanticCategory"] in hb_cats | |
| ] | |
| rows += [ | |
| { | |
| "id": f"mj:{r['id']}", | |
| "source": "multijail", | |
| "split": "harmful", | |
| "en": r["en"], | |
| "ar": r["ar"], | |
| "sw": r["sw"], | |
| } | |
| for r in mj | |
| if set(ast.literal_eval(r["tags"])) & mj_tags # a list literal in the CSV, never eval | |
| ] | |
| # near-miss benign: XSTest safe prompts whose focus word or note matches the category | |
| # lexicon | |
| for r in safe: | |
| blob = f"{r['focus']} {r['note']} {r['prompt']}".lower() | |
| if any(k in blob for k in kws): | |
| rows.append( | |
| { | |
| "id": f"xs:{r['id']}", | |
| "source": "xstest", | |
| "split": "benign", | |
| "en": r["prompt"], | |
| "xstest_type": r["type"], | |
| } | |
| ) | |
| with (OUT / f"pool_{cat}.jsonl").open("w") as f: | |
| for r in rows: | |
| f.write(json.dumps(r, ensure_ascii=False) + "\n") | |
| stats[cat] = { | |
| "harmful": sum(r["split"] == "harmful" for r in rows), | |
| "benign": sum(r["split"] == "benign" for r in rows), | |
| "with_human_ar_sw": sum("sw" in r for r in rows), | |
| } | |
| # generic benign control set: all XSTest safe prompts (over-refusal), plus the unsafe contrast | |
| # set | |
| with (OUT / "xstest_all.jsonl").open("w") as f: | |
| for r in xs: | |
| f.write( | |
| json.dumps( | |
| { | |
| "id": f"xs:{r['id']}", | |
| "type": r["type"], | |
| "label": r["label"], | |
| "en": r["prompt"], | |
| }, | |
| ensure_ascii=False, | |
| ) | |
| + "\n" | |
| ) | |
| (OUT / "build_stats.json").write_text(json.dumps(stats, indent=2)) | |
| print(json.dumps(stats, indent=2)) | |
| if __name__ == "__main__": | |
| main() | |