Datasets:
File size: 3,118 Bytes
af5b4af | 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 | """Assemble the public Hugging Face dataset folder.
Writes OUT/tasks.jsonl (one task per line: id, family, domain, level, an example instruction, the calibration screen stats and whether the task is in the submission), and
OUT/source/ (the environment, tools, tests, Dockerfile). The README card is kept by hand.
Usage: python tools/make_dataset.py --screen runs/a.jsonl runs/b.jsonl --selected runs/selected.txt --out dataset
"""
import argparse
import json
import os
import random
import shutil
import sys
from collections import defaultdict
sys.path.insert(0, ".")
from terminal_desk.families import FAMILIES, all_task_ids, instance_seed, parse_task_id # noqa: E402
SOURCE = ["terminal_desk", "tools", "tests", "Dockerfile", "openenv.yaml", "README.md", "LICENSE", "NOTICE",
".dockerignore", "vendor"]
def main() -> None:
p = argparse.ArgumentParser()
p.add_argument("--screen", required=True, nargs="+")
p.add_argument("--selected", required=True)
p.add_argument("--out", required=True)
a = p.parse_args()
stats = defaultdict(list)
for path in a.screen:
for line in open(path):
r = json.loads(line)
if "error" not in r:
stats[r["task_id"]].append(r)
selected = {ln.strip() for ln in open(a.selected) if ln.strip() and not ln.startswith("#")}
with open(os.path.join(a.out, "tasks.jsonl"), "w") as fh:
for tid in all_task_ids():
fid, level, _ = parse_task_id(tid)
fam = FAMILIES[fid]
built = fam.build(level, random.Random(instance_seed(tid))) # the fixed instance reset serves
rs = stats.get(tid, [])
rewards = [r["reward"] for r in rs]
fh.write(json.dumps({
"task_id": tid,
"family": fid,
"domain": fam.domain,
"level": level,
"title": fam.title,
"in_submission": tid in selected,
"screen_episodes": len(rs),
"screen_mean_reward": round(sum(rewards) / len(rewards), 3) if rewards else None,
"screen_full_solves": sum(1 for x in rewards if x == 1.0),
"screen_non_json_endings": sum(1 for r in rs if r["ended"] in ("invalid_reply", "truncated", "invalid_action")),
"screen_budget_endings": sum(1 for r in rs if r["ended"] in ("budget", "context")),
"screen_mean_steps": round(sum(r["steps"] for r in rs) / len(rs), 2) if rs else None,
"artifact_path": "/app/" + built.artifact.path,
"artifact_kind": built.artifact.kind,
"example_instruction": built.instruction,
}) + "\n")
src = os.path.join(a.out, "source")
shutil.rmtree(src, ignore_errors=True)
for name in SOURCE:
if os.path.isdir(name):
shutil.copytree(name, os.path.join(src, name), ignore=shutil.ignore_patterns("__pycache__", "qwen_tok"))
else:
shutil.copy(name, os.path.join(src, name))
print("dataset folder ready:", a.out)
if __name__ == "__main__":
main()
|