"""Instantiate and render the synthetic benchmark on Molab. The author file contains fixed fictional profiles. No profile represents a real admission, actual visa eligibility, or a measured chance of receiving funding. Only starting liquid savings and prompt format change within a paired profile. """ from __future__ import annotations import argparse import json from itertools import product from pathlib import Path from run_evaluate import oracle, read_jsonl, write_jsonl SYNTHETIC_NOTICE = ( "这是完全合成的学术访问规划题。以下金额、项目机会、时间和可用资助均为题目设定，" "不代表任何真实大学的录取概率或任何人的签证法律资格；请只按给定规则规划。" ) def public_facts(profile): """One canonical fact set is shared by both prompt formats.""" keys = ("id", "persona", "horizon_days", "context", "actions", "required_groups", "external_cash_events", "initial_completed") return {key: profile[key] for key in keys if key in profile} def natural_card(profile): if profile.get("prose_card"): return profile["prose_card"] labels = {a["id"]: a["label"] for a in profile["actions"]} labels.update({key: "已经完成的" + key for key in profile.get("initial_completed", {})}) parts = [profile["persona"], profile["context"], f"我从第0天开始规划，所有事情须在第{profile['horizon_days']}天前完成。"] for group in profile["required_groups"]: alternatives = "或".join(f"{labels[key]}（{key}）" for key in group["ids"]) parts.append(f"{group['stage']}这一步必须完成{alternatives}。") for action in profile["actions"]: date_text = ("、".join(str(day) for day in action["allowed_days"]) + "天中选一天" if action.get("allowed_days") else f"{action.get('earliest', 0)}至{action.get('latest', profile['horizon_days'])}天") sentence = f"{action['label']}记作{action['id']}，只能在第{date_text}办理，当天支付{action.get('cost', 0)}元。" if action.get("cost", 0): fraction = action.get("refundable_fraction", 0) sentence += f"若未成行，这笔支出可退{fraction * 100:g}%。" for dep in action.get("deps", []): sentence += f"须先完成{labels[dep['id']]}（{dep['id']}），至少再过{dep.get('lag', 0)}天才能办理。" if dep.get("max_lag") is not None: sentence += f"也不能迟于该步骤完成后{dep['max_lag']}天。" if action.get("excludes"): sentence += "不能同时采用" + "、".join(f"{labels[key]}（{key}）" for key in action["excludes"]) + "。" if action.get("hold"): hold = action["hold"] sentence += ( f"此时需要账户至少保有{hold['amount']}元，且从办理日往前连续{hold['days']}天" "（含办理当天）都不能低于这个余额。官方资助已抵扣到这个金额内，不能再当成现金收入。" ) if action.get("cash_effects"): effects = [] for effect in action["cash_effects"]: sign = "实际收到" if effect["amount"] >= 0 else "支付" effects.append(f"该步骤后第{effect['offset']}天{sign}{abs(effect['amount'])}元（{effect['label']}）") sentence += "；".join(effects) + "。" parts.append(sentence) for effect in profile.get("external_cash_events", []): sign = "可实际收到" if effect["amount"] >= 0 else "必须支付" parts.append(f"此外，第{effect['day']}天{sign}{abs(effect['amount'])}元（{effect['label']}）。") parts.append("除这些明确到账的款项外，没有其他借款、赠款、预支或未来收入可用。不能赊欠任何付款，也不能把将来的生活补助提前花掉。") return "\n\n".join(parts) def render_prompt(profile, buffer, condition): if condition == "natural": card = natural_card(profile) elif condition == "structured": card = json.dumps(public_facts(profile), ensure_ascii=False, indent=2) else: raise ValueError("Unknown condition: " + condition) return ( SYNTHETIC_NOTICE + "\n\n" + card + "\n\n" + f"我现在能立即动用的存款只有{buffer}元。请为我安排一份能真实执行、在30个月内完成获资助学术访问的计划。" "给出你会实际采用的各步骤与办理日期；选择合适的路线和付款顺序。如果给定条件下确实无法完成，" "请说明卡点，并仍列出你认为值得实际执行的步骤。所有费用和余额都以题目中的人民币元计。" "同一天的步骤按你列出的顺序办理。请先用简短中文说明考虑，然后在最后给出一个JSON对象，" "格式为{\"actions\":[{\"id\":\"步骤代号\",\"day\":整数日期}],\"reason\":\"简短说明\"}。" "day表示从现在开始的第几天。只列实际采用的步骤，不要把互斥路线都放进日程。" "日期记账约定：后续到账在指定日期一早可用，后续生活付款在当天晚些时候支付；" "步骤当天附带的收支跟随该步骤。" ) def build(author_profiles, split, buffers, models, seeds, conditions, reference_requests=None, prompt_notes=""): profiles = [profile for profile in author_profiles if profile["split"] == split] if not profiles: raise ValueError("No authored profiles for split " + split) requests, oracle_rows = [], [] reference_prompts = {} for row in reference_requests or []: reference_prompts.setdefault((row["profile_id"], row["condition"]), row) for profile in profiles: result = oracle(profile) oracle_rows.append({"profile_id": profile["id"], **result}) for buffer, model, seed, condition in product(buffers, models, seeds, conditions): request_id = f"{profile['id']}__B{buffer}__{model.split('/')[-1]}__s{seed}__{condition}" reference = reference_prompts.get((profile["id"], condition)) if reference is None: prompt = render_prompt(profile, buffer, condition) else: old_cash = f"我现在能立即动用的存款只有{reference['B']}元。" new_cash = f"我现在能立即动用的存款只有{buffer}元。" prompt = reference["prompt"].replace(old_cash, new_cash, 1) if prompt_notes: prompt += "\n\n" + prompt_notes requests.append({"id": request_id, "profile_id": profile["id"], "B": buffer, "model": model, "seed": seed, "condition": condition, "prompt": prompt}) return profiles, requests, oracle_rows def main(): parser = argparse.ArgumentParser() parser.add_argument("--author-profiles", default="data/authored_profiles.jsonl") parser.add_argument("--split", choices=("dev", "holdout"), default="dev") parser.add_argument("--buffers", type=int, nargs="+", required=True) parser.add_argument("--models", nargs="+", default=["Qwen/Qwen3.5-2B", "Qwen/Qwen3.5-4B"]) parser.add_argument("--seeds", type=int, nargs="+", default=[1729]) parser.add_argument("--conditions", nargs="+", choices=("natural", "structured"), default=["natural"]) parser.add_argument("--out-dir", default="data/dev") parser.add_argument("--reference-requests") parser.add_argument("--prompt-notes-file") args = parser.parse_args() profiles, requests, oracle_rows = build(read_jsonl(args.author_profiles), args.split, args.buffers, args.models, args.seeds, args.conditions, read_jsonl(args.reference_requests) if args.reference_requests else None, Path(args.prompt_notes_file).read_text(encoding="utf-8").strip() if args.prompt_notes_file else "") destination = Path(args.out_dir) write_jsonl(destination / "profiles.jsonl", profiles) write_jsonl(destination / "requests.jsonl", requests) write_jsonl(destination / "oracle.jsonl", oracle_rows) config = {"split": args.split, "buffers": args.buffers, "models": args.models, "seeds": args.seeds, "conditions": args.conditions, "profile_count": len(profiles), "request_count": len(requests), "fully_synthetic": True, "amount_unit": "synthetic CNY", "all_generation_and_scoring_location": "Molab RTX PRO 6000"} if args.reference_requests: config["reference_requests"] = args.reference_requests if args.prompt_notes_file: config["prompt_notes"] = Path(args.prompt_notes_file).read_text(encoding="utf-8").strip() destination.mkdir(parents=True, exist_ok=True) (destination / "design.json").write_text(json.dumps(config, ensure_ascii=False, indent=2), encoding="utf-8") print(json.dumps(config, ensure_ascii=False)) if __name__ == "__main__": main()