"""gen_hf2: real HF datasets -> alea rows, targeting JevBench-weak skills. Row schema (matches alea/dataset.py): {"state": str, "questions": {qid: {"type":..., "instructions":..., "criteria": {key: desc} | {"true":..., "false":...}}}, "targets": {qid: dist|float}, "family": str, "meta": {...}} choice: criteria = {label_key: option_text}; target = one-hot over keys order noul: instructions (+ true/false gloss); target = float(0/1) Datasets (real labeled data; NO JevBench items anywhere): clinc150 intent routing (fixes mode collapse) choice logiqa logical reasoning MCQ choice reclor LSAT-style logical reasoning choice math_qa GRE-style math MCQ choice gsm8k grade-school math -> choice over close numbers choice drop passage + numeric answers choice strategyqa implicit multi-hop yes/no noul creak commonsense fact verification noul bigbench date_understanding/logical_deduction/tracking/ web_of_lies/formal_fallacies/navigate/ temporal_sequences/disambiguation_qa choice/noul """ import argparse import json import random from datasets import load_dataset rng = random.Random(1234) def choice_row(state, instructions, options_texts, correct_idx, family, keys=None, meta=None): """options_texts: list of option strings; correct_idx into that list.""" n = len(options_texts) keys = keys or ["opt_%d" % i for i in range(n)] assert len(keys) == n and 0 <= correct_idx < n seen, final_opts, final_keys, new_idx = {}, [], [], None for i, t in enumerate(options_texts): t = str(t).strip() if not t: t = "(empty option %d)" % i if t in seen: if i == correct_idx: new_idx = seen[t] continue seen[t] = len(final_opts) final_opts.append(t) final_keys.append(keys[i]) if i == correct_idx: new_idx = len(final_opts) - 1 if new_idx is None or len(final_opts) < 2: return None tgt = [0.0] * len(final_opts) tgt[new_idx] = 1.0 return {"state": str(state)[:20000], "questions": {"q0": {"type": "choice", "instructions": instructions, "criteria": dict(zip(final_keys, final_opts))}}, "targets": {"q0": tgt}, "family": family, "meta": meta or {}} def noul_row(state, instructions, is_true, family, true_gloss="", false_gloss="", meta=None): crit = {} if true_gloss: crit["true"] = true_gloss if false_gloss: crit["false"] = false_gloss q = {"type": "noul", "instructions": instructions} if crit: q["criteria"] = crit return {"state": str(state)[:20000], "questions": {"q0": q}, "targets": {"q0": float(is_true)}, "family": family, "meta": meta or {}} def gen_clinc(n, rng): """Intent classification with distractor intents -> kills mode collapse.""" ds = load_dataset("clinc_oos", "plus", split="train") names = ds.features["intent"].names out = [] idxs = list(range(len(ds))) rng.shuffle(idxs) for i in idxs: if len(out) >= n: break r = ds[i] true_name = names[r["intent"]].replace("_", " ") k = rng.randint(3, 6) others = rng.sample([x for x in names if x != names[r["intent"]]], k) opts = [true_name] + [o.replace("_", " ") for o in others] order = list(range(len(opts))) rng.shuffle(order) shuffled = [opts[j] for j in order] ci = shuffled.index(true_name) st = ("Customer message received by the virtual assistant:\n\n\"" + r["text"].strip() + "\"") row = choice_row(st, "Classify the user's intent into exactly one of " "the listed intents.", shuffled, ci, "hf_clinc_intent", meta={"src": "clinc150"}) if row: out.append(row) return out def gen_logiqa(n, rng): ds = None for nm in ("jeggers/logiqa", "EleutherAI/logiqa", "lucasmccabe/logiqa"): try: ds = load_dataset(nm, split="train") break except Exception: continue if ds is None: return [] out = [] for r in ds: if len(out) >= n: break st = "Passage:\n%s\n\nQuestion:\n%s" % (r["context"].strip(), r["query"].strip()) opts = [o.strip() for o in r["options"]] row = choice_row(st, "Choose the option that is logically entailed or " "best answers the question.", opts, int(r["correct_option"]), "hf_logiqa", meta={"src": "logiqa"}) if row: out.append(row) return out def gen_reclor(n, rng): ds = load_dataset("metaeval/reclor", split="train") out = [] for r in ds: if len(out) >= n: break st = "Passage:\n%s\n\nQuestion:\n%s" % (r["context"].strip(), r["question"].strip()) opts = [o.strip() for o in r["answers"]] row = choice_row(st, "Choose the option that best completes the " "logical argument.", opts, int(r["label"]), "hf_reclor", meta={"src": "reclor"}) if row: out.append(row) return out def gen_mathqa(n, rng): ds = load_dataset("allenai/math_qa", split="train", trust_remote_code=True) out = [] letter = "abcde" for r in ds: if len(out) >= n: break opts = [o.strip() for o in r["options"].split(" , ")] if len(opts) != 5: continue ci = letter.find(r["correct"]) if ci < 0: continue st = "Problem:\n%s" % r["Problem"].strip() row = choice_row(st, "Solve the problem and choose the correct " "answer.", opts, ci, "hf_mathqa", meta={"src": "mathqa"}) if row: out.append(row) return out def gen_gsm8k(n, rng): ds = load_dataset("openai/gsm8k", "main", split="train") out = [] fmt = lambda x: ("%g" % x) for r in ds: if len(out) >= n: break try: ans = float(r["answer"].split("####")[-1].replace(",", "").strip()) except ValueError: continue cands = {ans * f for f in (0.5, 0.75, 0.9, 1.1, 1.25, 1.5, 2.0)} cands |= {ans + d for d in (-3, -2, -1, 1, 2, 3)} cands = [c for c in cands if abs(c - ans) > max(0.51, 0.04 * abs(ans))] rng.shuffle(cands) opts = [fmt(x) for x in ([ans] + cands[:3])] order = list(range(4)); rng.shuffle(order) shuffled = [opts[j] for j in order] ci = shuffled.index(fmt(ans)) st = "Problem:\n%s" % r["question"].strip() row = choice_row(st, "Solve the problem. Choose the exact final " "answer.", shuffled, ci, "hf_gsm8k", meta={"src": "gsm8k"}) if row: out.append(row) return out def gen_drop(n, rng): ds = load_dataset("ucinlp/drop", split="train") out = [] for r in ds: if len(out) >= n: break spans = r["answers_spans"]["spans"] if len(spans) != 1: continue ans = spans[0].strip() if not ans: continue try: a = float(ans.replace(",", "").replace("$", "").replace("%", "")) cands = {a * f for f in (0.5, 0.75, 1.25, 1.5, 2.0)} | {a + d for d in (-2, -1, 1, 2)} cands = [c for c in cands if abs(c - a) > max(0.4, 0.03 * abs(a))] rng.shuffle(cands) fmt = lambda x: ("%g" % x) opts = [fmt(x) for x in ([a] + cands[:3])] order = list(range(4)); rng.shuffle(order) shuffled = [opts[j] for j in order] ci = shuffled.index(fmt(a)) st = "Passage:\n%s\n\nQuestion:\n%s" % (r["passage"][:6000], r["question"]) row = choice_row(st, "Answer the question about the passage with " "the correct number.", shuffled, ci, "hf_drop", meta={"src": "drop"}) if row: out.append(row) except ValueError: if ans.lower() in ("yes", "no"): st = "Passage:\n%s\n\nQuestion:\n%s" % (r["passage"][:6000], r["question"]) out.append(noul_row(st, "Answer yes or no based only on the " "passage.", ans.lower() == "yes", "hf_drop_noul", meta={"src": "drop"})) return out def gen_strategyqa(n, rng): ds = None for name in ("ChilleD/StrategyQA", "tasksource/strategy-qa"): try: ds = load_dataset(name, split="train") break except Exception: continue if ds is None: return [] out = [] for r in ds: if len(out) >= n: break q = r.get("question") or r.get("Question") or "" a = r.get("answer", r.get("Answer")) if not q or a is None: continue out.append(noul_row("Question: %s" % q.strip(), "Answer yes or no. Reason about the implied " "facts carefully.", bool(a), "hf_strategyqa", meta={"src": "strategyqa"})) return out def gen_creak(n, rng): ds = None for name in ("amariucaitheodor/creak", "creak"): try: ds = load_dataset(name, split="train") break except Exception: continue if ds is None: return [] out = [] for r in ds: if len(out) >= n: break sent = r.get("sentence", "") lab = r.get("label", "") if not sent or lab not in ("true", "false"): continue out.append(noul_row("Claim: %s" % sent.strip(), "Decide whether the claim is factually true.", lab == "true", "hf_creak", meta={"src": "creak"})) return out BB_TASKS = [ ("date_understanding", "choice"), ("logical_deduction", "choice"), ("tracking_shuffled_objects", "choice"), ("formal_fallacies_syllogisms_negation", "noul"), ("navigate", "noul"), ("temporal_sequences", "choice"), ("disambiguation_qa", "choice"), ("logical_fallacy_detection", "choice"), ("logic_grid_puzzle", "choice"), ("logical_sequence", "choice"), ("logical_args", "choice"), ("repeat_copy_logic", "choice"), ("chess_state_tracking", "choice"), ] def gen_bigbench(n_per_task, rng): out = [] for task, kind in BB_TASKS: try: ds = None for sp in ("train", "validation"): try: ds = load_dataset("tasksource/bigbench", task, split=sp) break except Exception: continue except Exception as e: print("bigbench skip", task, str(e)[:60]) continue cnt = 0 idxs = list(range(len(ds))) rng.shuffle(idxs) for i in idxs: if cnt >= n_per_task: break r = ds[i] inp = (r.get("inputs") or r.get("input") or "").strip() import ast as _ast tgts = r.get("multiple_choice_targets") or [] scrs = r.get("multiple_choice_scores") or [] if isinstance(tgts, str): try: tgts = _ast.literal_eval(tgts) except Exception: tgts = [] if isinstance(scrs, str): try: scrs = _ast.literal_eval(scrs) except Exception: scrs = [] if not inp or len(tgts) < 2 or not any(scrs): continue ci = list(scrs).index(1) opts = [str(o).strip() for o in tgts] if kind == "noul" and len(opts) == 2: yesish = [o.lower().startswith(("yes", "true", "valid", "correct")) for o in opts] if sum(yesish) == 1: out.append(noul_row(inp, "Answer yes or no.", ci == yesish.index(True), "hf_bb_" + task, meta={"src": task})) cnt += 1 continue row = choice_row(inp, "Choose the correct answer.", opts, ci, "hf_bb_" + task, meta={"src": task}) if row: out.append(row) cnt += 1 print("bb", task, cnt) return out GENS = { "clinc": gen_clinc, "logiqa": gen_logiqa, "reclor": gen_reclor, "gsm8k": gen_gsm8k, "drop": gen_drop, "strategyqa": gen_strategyqa, "bigbench": gen_bigbench, } def main(): ap = argparse.ArgumentParser() ap.add_argument("--out", default="data/hf2.jsonl") ap.add_argument("--per", type=int, default=6000) ap.add_argument("--bb-per", type=int, default=3000) ap.add_argument("--only", default="") args = ap.parse_args() rows = [] only = set(args.only.split(",")) if args.only else None for name, fn in GENS.items(): if only and name not in only: continue try: got = fn(args.bb_per if name == "bigbench" else args.per, rng) except Exception as e: print(name, "FAILED", type(e).__name__, str(e)[:120]) continue print(name, len(got)) rows += got rng.shuffle(rows) with open(args.out, "w") as f: for r in rows: f.write(json.dumps(r, ensure_ascii=False) + "\n") print("TOTAL", len(rows), "->", args.out) if __name__ == "__main__": main()