"""gen_jb3: LONG-document reasoning — multi-section policies with nested exceptions, judge-style verdicts, multi-constraint chains. Hard tier needs 1-4k token docs where the answer depends on composing several sections in order (base rule -> endorsement override -> exception-to-endorsement -> global condition). These generators build that structure deterministically so the correct answer is exact. """ import argparse import json import random rng = random.Random(31337) def choice(state, instr, crit, correct_key, family): keys = list(crit.keys()) tgt = [0.0] * len(keys) tgt[keys.index(correct_key)] = 1.0 return {"state": state, "questions": {"q0": {"type": "choice", "instructions": instr, "criteria": crit}}, "targets": {"q0": tgt}, "family": family} def noul(state, instr, val, family): return {"state": state, "questions": {"q0": {"type": "noul", "instructions": instr}}, "targets": {"q0": float(val)}, "family": family} # ---------- long insurance/loan/policy docs ---------- def gen_longpolicy(n): """Multi-section policy with exception chains -> adjudication.""" out = [] for _ in range(n): # --- document skeleton --- cov_window = rng.choice([30, 60, 90]) late_days = rng.choice([30, 45]) ded = rng.choice([250, 500]) cap = rng.choice([2000, 5000, 10000]) region_ok = rng.choice([True, False]) secs = [ "MASTER SERVICES AND PROTECTION AGREEMENT", "Document ref: MPA-%d" % rng.randint(1000, 9999), "In force: %s through %s." % ( rng.choice(["01 Jan 2024", "01 Feb 2024"]), rng.choice(["31 Dec 2025", "31 Dec 2026"])), "", "1. DEFINITIONS", " 1.1 'Covered Event' means an event listed in Section 3.", " 1.2 'Territory' means the coverage area in Schedule A.", " 1.3 'Prompt Notice' means notice within %d days." % late_days, "", "2. SCOPE", " 2.1 This agreement applies to claims arising during the " "in-force period.", " 2.2 Claims arising outside the Territory are excluded " "absent a Territory Endorsement.", "", "3. COVERED EVENTS", " 3.1 Accidental damage reported within %d days." % cov_window, " 3.2 Theft with a filed police report.", " 3.3 Mechanical failure not caused by misuse.", "", "4. EXCLUSIONS", " 4.1 Cosmetic damage is excluded.", " 4.2 Events arising from unauthorized modification are " "excluded.", " 4.3 Events occurring before the in-force date are excluded.", "", "5. CONDITIONS PRECEDENT", " 5.1 Prompt Notice is a condition precedent to coverage.", " 5.2 A deductible of $%d applies per claim." % ded, "", "6. ENDORSEMENTS", ] # endorsements create the exception chains e_theft_no_police = rng.random() < 0.4 e_territory = rng.random() < 0.5 e_mod = rng.random() < 0.35 if e_theft_no_police: secs.append(" E-T1: theft claims without a police report are " "covered up to $%d if reported with two forms of " "proof of ownership." % cap) if e_territory: secs.append(" E-TR: the Territory exclusion in 2.2 is waived " "for losses occurring within 50 km of the Territory " "boundary.") if e_mod: secs.append(" E-M1: Section 4.2 does not apply where the " "modification was performed by an authorized " "service center.") secs += ["", "7. CLAIMS PROCEDURE", " 7.1 All claims must be filed via the portal.", " 7.2 The insurer responds within 15 business days.", "", "8. GENERAL", " 8.1 This agreement is governed by the laws of the " "issuing jurisdiction.", " 8.2 Nothing herein limits statutory rights."] # --- the claim --- event = rng.choice(["accident", "theft", "mechanical"]) reported = rng.choice(["prompt", "late"]) outside = rng.random() < 0.4 police = rng.random() < 0.6 modded = rng.random() < 0.35 auth_mod = modded and (rng.random() < 0.5) facts = ["Loss date: %s 2025 (in-force period)." % rng.choice(["Apr 3", "Jul 19", "Oct 8"]), "Event type: %s." % event] if event == "accident": facts.append("Damage is structural, not cosmetic.") facts.append("Reported %d days after the event." % (rng.randint(1, cov_window - 3) if reported == "prompt" else cov_window + rng.randint(3, 20))) if event == "theft": facts.append("Police report %s." % ("filed same day" if police else "was NOT filed")) if e_theft_no_police and not police: facts.append("Claimant supplied %s." % rng.choice(["two proofs of ownership", "one proof of ownership"])) if event == "mechanical": facts.append("Unit %s." % ("was modified by an authorized service center" if auth_mod else "has an unauthorized third-party modification" if modded else "is unmodified")) facts.append("Reported %d days after the event." % rng.randint(1, cov_window)) if outside: facts.append("Loss occurred %d km outside the Territory " "boundary." % rng.choice([20, 80])) if reported == "late" and event != "accident": facts.append("Claim filed %d days after the loss." % (late_days + rng.randint(1, 15))) elif event != "accident": facts.append("Claim filed %d days after the loss." % rng.randint(1, late_days - 2)) # --- adjudicate --- if event == "accident" and reported == "late": correct = "deny_late" elif event == "accident": correct = "pay" elif event == "theft": if police: correct = "pay" elif e_theft_no_police: two_proofs = any("two proofs" in f for f in facts) correct = "pay_capped" if two_proofs else "deny_condition" else: correct = "deny_condition" else: # mechanical if modded and auth_mod and e_mod: correct = "pay" elif modded and auth_mod and not e_mod: correct = "deny_excluded" elif modded and not auth_mod: correct = "deny_excluded" else: correct = "pay" if correct == "pay" and outside: near = "20 km" in " ".join(facts) correct = "pay" if (e_territory and near) else "deny_excluded" elif outside and correct.startswith("pay"): near = "20 km" in " ".join(facts) correct = "pay" if (e_territory and near) else "deny_excluded" if reported == "late" and event != "accident" and correct == "pay": correct = "deny_late" crit = { "pay": "Pay the claim less the deductible", "pay_capped": "Pay subject to the endorsement cap", "deny_late": "Deny: notice condition not met", "deny_excluded": "Deny: exclusion applies", "deny_condition": "Deny: a coverage condition was not met", } st = "\n".join(secs) + "\n\nCLAIM:\n" + "\n".join(facts) out.append(choice(st, "Adjudicate the claim. Apply sections in " "order; endorsements modify earlier sections; " "conditions precedent bind first.", crit, correct, "jb3_longpolicy")) return out # ---------- judge: verify a verdict ---------- def gen_judge(n): """Doc + candidate verdict -> agree/disagree (judge_hard style).""" out = [] for _ in range(n): a = rng.randint(100, 900) b = rng.randint(100, 900) c = rng.randint(100, 900) col = rng.choice(["revenue", "cost", "units"]) rec = {"revenue": a, "cost": b, "units": c} rec["margin"] = round((a - b) / a, 3) st = ("Record:\n revenue: %d\n cost: %d\n units: %d\n margin: " "%.3f\n region: %s\n quarter: Q%d" % (a, b, c, rec["margin"], rng.choice(["EU", "NA", "APAC"]), rng.randint(1, 4))) # candidate verdict — correct half the time kind = rng.choice(["margin", "sum", "max"]) if kind == "margin": true_v = rec["margin"] wrong_v = round((a - b) / b, 3) claim = "the margin is %.3f" % (true_v if rng.random() < 0.5 else wrong_v) ok = ("%.3f" % true_v) in claim elif kind == "sum": true_v = a + b wrong_v = a + b + rng.choice([-1, 1]) * rng.randint(5, 50) claim = "revenue plus cost equals %d" % (true_v if rng.random() < 0.5 else wrong_v) ok = str(true_v) in claim else: big = max(rec.values()) wrong = min(rec.values()) claim = "the largest field is %d" % (big if rng.random() < 0.5 else wrong) ok = str(big) in claim st += "\n\nAnalyst verdict: \"%s\"." % claim out.append(choice(st, "Judge the verdict: does it follow from the " "record?", {"correct": "Verdict is correct", "incorrect": "Verdict is incorrect"}, "correct" if ok else "incorrect", "jb3_judge")) return out # ---------- multi-hop temporal chains ---------- def gen_temporal_chain(n): """3-4 step date/shift chains — multi_hop style.""" import datetime as dt out = [] for _ in range(n): y = rng.choice([2024, 2025]) m, d = rng.randint(1, 12), rng.randint(1, 28) try: base = dt.date(y, m, d) except ValueError: continue hops = [rng.randint(3, 14) for _ in range(3)] wds = [h * rng.choice([1, 1]) for h in hops] steps_txt = [] cur = base for i, h in enumerate(hops): steps_txt.append("Step %d: add %d days." % (i + 1, h)) cur += dt.timedelta(days=h) # ask for final weekday or date ask = rng.choice(["weekday", "date"]) st = ("Start date: %s.\n%s\n" % (base.strftime("%A, %B %d, %Y"), "\n".join(steps_txt))) if ask == "weekday": opts = ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"] rng.shuffle(opts) crit = {o: "final day is %s" % o for o in opts} out.append(choice(st, "Apply each step in order and choose the " "final weekday.", crit, cur.strftime("%A"), "jb3_temporal_chain")) else: opts = [cur.strftime("%B %d")] for _ in range(3): w = cur + dt.timedelta(days=rng.choice([-3, -2, -1, 1, 2, 3])) opts.append(w.strftime("%B %d")) order = list(range(4)); rng.shuffle(order) shuffled = [opts[j] for j in order] crit = {"d%d" % i: shuffled[i] for i in range(4)} out.append(choice(st, "Apply each step in order and choose the " "final date.", crit, "d%d" % order.index(0), "jb3_temporal_chain")) return out GENS = {"longpolicy": gen_longpolicy, "judge": gen_judge, "temporal_chain": gen_temporal_chain} def main(): ap = argparse.ArgumentParser() ap.add_argument("--out", default="data/jb3.jsonl") ap.add_argument("--per", type=int, default=8000) args = ap.parse_args() rows = [] for name, fn in GENS.items(): got = fn(args.per) 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()