Datasets:
File size: 5,399 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 | """Build the frozen post-offer main benchmark on Molab before model inference."""
from __future__ import annotations
import json
from itertools import product
from pathlib import Path
from build_core_dev import core_profile, core_prompt
from run_evaluate import oracle, read_jsonl, write_jsonl
BUFFERS = [5000, 7000, 8000, 9000, 10000, 11000, 12000, 13000, 14000, 15000,
16000, 17000, 18000, 20000, 22000, 24000, 26000, 28000, 40000]
SEEDS = [1729, 2718]
MODELS = ["Qwen/Qwen3.5-2B", "Qwen/Qwen3.5-4B"]
def main():
destination = Path("data/main_v1")
original = [profile for profile in read_jsonl("data/authored_profiles.jsonl")
if profile["split"] == "holdout"]
profiles = [core_profile(profile) for profile in original]
references, requests = [], []
for profile in profiles:
result = oracle(profile)
references.append({"profile_id": profile["id"], **result})
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 / "original_profiles.jsonl", original)
write_jsonl(destination / "profiles.jsonl", profiles)
write_jsonl(destination / "oracle.jsonl", references)
write_jsonl(destination / "requests.jsonl", requests)
coverage = [{"profile_id": reference["profile_id"],
"minimum_initial_cash": reference["minimum_initial_cash"],
"routes": [{"route_id": route["route_id"],
"minimum_initial_cash": route["minimum_initial_cash"],
"has_grid_below": any(buffer < route["minimum_initial_cash"] for buffer in BUFFERS),
"has_grid_at_or_above": any(buffer >= route["minimum_initial_cash"] for buffer in BUFFERS)}
for route in reference["route_minima"]]}
for reference in references]
design = {
"version": "main_v1", "task": "conditional post-offer academic visit execution",
"profile_ids": [profile["id"] for profile in profiles], "profile_count": len(profiles),
"buffers": BUFFERS, "models": MODELS, "seeds": SEEDS, "conditions": ["natural"],
"world": "W0", "request_count": len(requests), "fully_synthetic": True,
"amount_unit": "synthetic CNY", "initial_completed": {"O": 0},
"output_max_tokens_both_models": 32768, "context_length_both_models": 131072,
"thinking": {"Qwen/Qwen3.5-2B": "off native", "Qwen/Qwen3.5-4B": "on native"},
"sampling": "each model's separately verified official native-mode recommendation",
"source_profiles": "The 12 fixed original holdout profiles authored before any holdout model outputs",
"prompt_template": "core_prompt from the single dev_v3 post-offer diagnostic, unchanged",
"buffer_selection": "Frozen grid selected from deterministic available-route reserves, without holdout model outputs",
"all_generation_and_scoring_location": "Molab RTX PRO 6000",
"route_grid_coverage": coverage,
}
destination.mkdir(parents=True, exist_ok=True)
(destination / "design.json").write_text(json.dumps(design, ensure_ascii=False, indent=2), encoding="utf-8")
(destination / "change_note.txt").write_text(
"Main_v1 adopts the conditional post-offer scope and exact core_prompt template from the one dev_v3 diagnostic. "
"All 12 previously fixed holdout profiles are mechanically transformed by the same core_profile function. "
"O is an already-completed day0 initial fact. Past preparation actions and already-paid fees are removed; "
"B is current remaining liquid cash. Future absolute allowed dates and the horizon are shifted by original O day. "
"Future costs, amounts, receipt offsets, refunds, holds, dependency lags and exclusions remain unchanged. "
"Recognized official sponsorship has already reduced residual proof amounts; full sponsorship removes the cash "
"hold rather than imposing an invented full bank balance requirement. The funding routes and every permitted "
"calendar/order are re-enumerated remotely. Original and transformed profiles, references and full requests are saved. "
"Both native systems use seeds1729/2718 and common32768 output cap/131072 context. The frozen19-buffer grid "
"was selected from pure reference thresholds, not model success transitions. No holdout model output informed "
"the facts, grid or seed choice, and none may be adjusted after main inference. Multiple scope/interface/budget "
"changes from earlier development versions cannot be attributed to a single cause; all those results remain separate. "
"The dev_v3 4B success on every reference-feasible diagnostic is contrary evidence to a universal liquidity-cliff claim.\n",
encoding="utf-8")
print(json.dumps({"request_count": len(requests), "profile_count": len(profiles),
"route_grid_coverage": coverage}, ensure_ascii=False), flush=True)
if __name__ == "__main__":
main()
|