Datasets:
File size: 11,633 Bytes
557f178 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 | """Build the one prespecified after-offer diagnostic on Molab only."""
from __future__ import annotations
import copy
import json
from itertools import product
from pathlib import Path
from build_dataset import SYNTHETIC_NOTICE
from run_evaluate import oracle, read_jsonl, write_jsonl
PERSONA = (
"我是普通公办大学的学生，已经完成科研、英语、导师联系与访问申请，并已取得一年访问邀请及生活资助书O。"
"此前准备费用已经支付。今天是取得确认后的第0天，我现在要办好剩余手续、实际出发并完成一年访问。"
"正常生活已有保障，题目给的B是今天实际可立即动用的剩余现金；家里不能追加，也没有借款渠道。"
)
def core_profile(source):
profile = copy.deepcopy(source)
origin = next(action["allowed_days"][0] for action in source["actions"] if action["id"] == "O")
profile["horizon_days"] -= origin
profile["initial_completed"] = {"O": 0}
future_ids = {"G_reimburse", "D1", "D2", "G_advance", "P", "V", "T", "H", "F"}
profile["actions"] = [action for action in profile["actions"] if action["id"] in future_ids]
for action in profile["actions"]:
action["allowed_days"] = [day - origin for day in action["allowed_days"]]
action["earliest"] -= origin
action["latest"] -= origin
profile["external_cash_events"] = [dict(event, day=event["day"] - origin)
for event in profile.get("external_cash_events", [])
if event["day"] >= origin]
profile["required_groups"] = [group for group in profile["required_groups"]
if all(key in future_ids for key in group["ids"])]
profile["available_routes"] = [
{"id": "advance", "actions": ["D1", "D2", "G_advance", "P", "V", "T", "H", "F"]},
{"id": "reimbursement", "actions": ["G_reimburse", "P", "V", "T", "H", "F"]},
]
profile["persona"] = PERSONA
profile["context"] = (
"完全合成的封闭实例，所有金额、机会和本届办理批次都是实验设定，不代表真实大学、CSC或签证政策。"
"邀请及资助书O是第0天已完成的初始事实，无需重新办理或列入未来日程；资助书不等于已到账现金。"
"过去科研、语言和申请费用已支付，当前B是支付之后的实际余额，不再扣除那些历史费用。"
)
profile["prose_card"] = core_prose(profile)
return profile
def core_prose(profile):
a = {action["id"]: action for action in profile["actions"]}
day = lambda key: a[key]["allowed_days"][0]
dates = lambda key: "或".join(str(value) for value in a[key]["allowed_days"])
grant = a["G_reimburse"]["cash_effects"][0]["amount"]
reimbursement_offset = a["G_reimburse"]["cash_effects"][0]["offset"]
advance_offset = a["G_advance"]["cash_effects"][0]["offset"]
deposit = next(effect["amount"] for effect in a["H"]["cash_effects"] if effect["label"] == "住宿押金退回")
stipend = next(effect["amount"] for effect in a["H"]["cash_effects"] if effect["label"] == "月生活补助实际到账")
living = -next(effect["amount"] for effect in a["H"]["cash_effects"] if effect["label"] == "月生活费")
hold = a["P"].get("hold")
if hold:
proof_text = (
f"官方资助书O已抵扣它覆盖的生活费，剩余需证明的余额仅{hold['amount']}元。"
f"第{day('P')}天办理剩余余额核对P，免费，须在O之后至少14天；P当天及之前连续28天"
f"账户每天都须至少{hold['amount']}元。"
)
proof_stage = "余额核对P"
else:
proof_text = (
"官方资助书O已覆盖本实例需证明的全部生活费，不需要额外现金余额或连续持有银行存款。"
f"第{day('P')}天核对这份官方资助书P，免费，须在O之后至少14天。"
)
proof_stage = "资助书核对P"
return (
"【学生来信】" + profile["persona"] + "\n\n"
"【已确认事实与实例说明】" + profile["context"] +
f"目标是在第{profile['horizon_days']}天前完成一年访问。以下所有日期以今天第0天为原点，"
"是本届仅有的办理批次，不接受其他日期。\n\n"
f"【出行资助邮件】差旅支持有两种互斥方式，金额同为{grant}元，必须选择一种。"
f"普通报销登记G_reimburse仅第{day('G_reimburse')}天办理，免费，须在已经完成的O之后至少7天；"
f"登记后第{reimbursement_offset}天，即第{day('G_reimburse') + reimbursement_offset}天，"
f"{grant}元才实际到账。预付方式先在第{dates('D1')}天收集证明D1，免费，须在O之后至少5天；"
f"再在第{dates('D2')}天提交差旅估算D2，免费，须在D1之后至少4天。"
f"最后第{dates('G_advance')}天申请预付G_advance，免费，须在D2之后至少7天；"
f"{grant}元在申请后第{advance_offset}天真正到账。G_reimburse和G_advance不能并用。\n\n"
"【行前行政通知】" + proof_text + "这是合成行政核对，不认定真实签证资格。"
f"第{day('V')}天办理行前手续V，支付{a['V']['cost']}元不退，须在P之后至少1天。"
f"第{day('T')}天预订交通T，支付{a['T']['cost']}元，其中50%可退，须在V之后至少7天。\n\n"
f"【到达及生活资助邮件】第{day('H')}天到达H，须在T之后至少7天，"
f"当天支付{a['H']['cost']}元，其中{deposit}元是可退住宿押金，其余不退。"
"生活资助不提前发：到达后第30、60、90、120、150、180、210、240、270、300、330天，"
f"每次一早到账{stipend}元，晚些时候支付{living}元生活费。到达后第360天退回{deposit}元押金。"
f"第{day('F')}天完成访问F，免费，须在H之后至少365天。\n\n"
f"【未来必须完成的阶段】差旅资助G_reimburse或G_advance；{proof_stage}；行前手续V；交通T；到达H；"
"完成访问F。预付方式的免费前置手续D1和D2也须按通知完成。除明确到账的这些款项外，"
"没有其他收入、赠款或借款；已确认的资助书本身不能当作可花现金。"
)
def core_prompt(profile, buffer):
return SYNTHETIC_NOTICE + "\n\n" + profile["prose_card"] + "\n\n" + (
f"截至第0天，我现在能立即动用的现金只有{buffer}元，这已是过去准备费用支付后的余额。"
"请仅规划未来步骤，安排能真实执行并在本实例给定期限内完成一年访问的计划。"
"给出实际采用的步骤代号与办理日期，选择融资路线和付款顺序。若确实无法完成，说明卡点，"
"并仍列出当前值得且能够实际执行的步骤。日期只能从通知列出的本届办理批次中选择，"
"满足最短间隔后也不能自行改变日期。逐项列出所选方案全部未来必需步骤，包括免费的前置手续和核对；"
"不能改名、合并或默认自动完成。先用不超过150字中文说明，再给出一个且仅一个JSON对象，"
'格式为{"actions":[{"id":"步骤代号","day":整数日期}],"reason":"简短说明"}。'
"day从今天第0天起算，同日步骤按列出的顺序办理。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 main():
destination = Path("data/dev_v3")
sources = [profile for profile in read_jsonl("data/dev_v1/profiles.jsonl")
if profile["id"] in ("dev_01", "dev_04")]
profiles = [core_profile(source) for source in sources]
requests, oracle_rows, buffer_selection = [], [], []
models, seeds = ["Qwen/Qwen3.5-2B", "Qwen/Qwen3.5-4B"], [1729]
for profile in profiles:
result = oracle(profile)
oracle_rows.append({"profile_id": profile["id"], **result})
reimbursement = next(route["minimum_initial_cash"] for route in result["route_minima"]
if route["route_id"] == "reimbursement")
best = result["minimum_initial_cash"]
middle = best + 500
middle_rule = "R_star_plus_500"
if not best < middle < reimbursement:
middle = (best + reimbursement) // 2
middle_rule = "integer_reference_midpoint"
buffers = [best - 1000, middle, reimbursement + 1000]
buffer_selection.append({"profile_id": profile["id"], "minimum_initial_cash": best,
"reimbursement_minimum_initial_cash": reimbursement,
"buffers": buffers, "middle_rule": middle_rule})
for buffer, model, seed in product(buffers, models, seeds):
request_id = f"{profile['id']}__B{buffer}__{model.split('/')[-1]}__s{seed}__natural"
requests.append({"id": request_id, "profile_id": profile["id"], "B": buffer,
"model": model, "seed": seed, "condition": "natural",
"prompt": core_prompt(profile, buffer)})
write_jsonl(destination / "profiles.jsonl", profiles)
write_jsonl(destination / "requests.jsonl", requests)
write_jsonl(destination / "oracle.jsonl", oracle_rows)
destination.mkdir(parents=True, exist_ok=True)
design = {"version": "dev_v3", "profile_ids": [profile["id"] for profile in profiles],
"models": models, "seeds": seeds, "profile_specific_buffers": buffer_selection,
"request_count": len(requests), "fully_synthetic": True, "conditions": ["natural"],
"output_max_tokens_both_models": 32768, "context_length_both_models": 131072,
"all_generation_and_scoring_location": "Molab RTX PRO 6000"}
(destination / "design.json").write_text(json.dumps(design, ensure_ascii=False, indent=2), encoding="utf-8")
(destination / "change_note.txt").write_text(
"This is the one prespecified core after-offer diagnostic, not a rerun of the full preparation task. "
"Sources are the unchanged authored dev_01 and dev_04 facts. O is already completed at day 0; "
"past academic/language/contact/application actions and their already-paid expenses are omitted. "
"Initial B is remaining spendable cash after those past payments. Future absolute dates and horizon "
"are shifted by the original O day. Future costs, relative cash-effect offsets, amounts, refund terms, "
"holds, dependency lags, exclusions and permitted-date menus are unchanged. Initial completed O0 "
"satisfies its published prerequisite lags. Only two future funding routes remain. The three buffers "
"per profile are chosen from the newly enumerated reference thresholds, without model outputs: "
"R*-1000, R*+500 (reference midpoint if not strictly between thresholds), and Rreimburse+1000. "
"Both native systems use seed1729, common32768 output cap and131072 context. Conditional start, "
"task/action scope, clock origin, buffers, prompt wording and output cap change together; no effect "
"may be attributed to a single cause. All previous dev results remain separate. No money or rule "
"was changed to manufacture a result.\n", encoding="utf-8")
print(json.dumps(design, ensure_ascii=False), flush=True)
if __name__ == "__main__":
main()
|