File size: 4,888 Bytes
81663e8 | 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 | """Materialize the fixed train/test tasks from corrected, pinned Harbor datasets."""
from __future__ import annotations
import argparse
import hashlib
import json
import shutil
from pathlib import Path
import tomllib
ROOT = Path(__file__).resolve().parent
def task_rows(split):
return json.loads((ROOT / "data" / f"{split}.json").read_text())
def write_json(path, data):
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(data, indent=2) + "\n")
DATASETS = {
"train": (
"FineEnvs/SmolDataEnvs-harbor-train",
"4719635555de1666f374d847baebb500d368e493",
),
"test": (
"FineEnvs/SmolDataEnvs-harbor-test",
"b130595b579fc6026ad5762925a9a3b40e72f5bb",
),
}
GRADER_SHA256 = "16f906e9248fa8b7e1141939e425e88e59d52d65e98e360bbacef3720a43629f"
def prepare(destination, sources=None):
import tomli_w
from huggingface_hub import snapshot_download
destination = Path(destination).resolve()
manifests = {}
for split in ("train", "test"):
rows = task_rows(split)
repo, revision = DATASETS[split]
source = (sources or {}).get(split)
if source is None:
source = snapshot_download(
repo,
repo_type="dataset",
revision=revision,
allow_patterns=[f"tasks/{r['name']}/*" for r in rows],
max_workers=16,
)
source = Path(source)
tasks = []
for row in rows:
folder = source / "tasks" / row["name"]
instruction = (folder / "instruction.md").read_bytes()
if hashlib.sha256(instruction).hexdigest() != row["instruction_sha256"]:
raise ValueError(
f"Historical task instruction changed: {split}/{row['name']}"
)
grader = (folder / "tests/grader.py").read_bytes()
if hashlib.sha256(grader).hexdigest() != GRADER_SHA256:
raise ValueError(
f"Expected corrected numeric-only Harbor grader: {folder}"
)
target = destination / "datasets" / split / "tasks" / row["name"]
target.mkdir(parents=True, exist_ok=True)
for rel in (
"instruction.md",
"task.toml",
"environment/Dockerfile",
"environment/pull_bucket.py",
"tests/grader.py",
"tests/test.sh",
):
dst = target / rel
dst.parent.mkdir(parents=True, exist_ok=True)
shutil.copyfile(folder / rel, dst)
spec = tomllib.loads((target / "task.toml").read_text())
spec["environment"]["memory_mb"] = 4096
spec["environment"]["cpus"] = 1
# Cluster user IDs may not exist inside Daytona's user namespace.
spec["environment"].setdefault("env", {})["TAR_OPTIONS"] = "--no-same-owner"
(target / "task.toml").write_text(tomli_w.dumps(spec))
hashes = {
str(f.relative_to(destination / "datasets" / split)): hashlib.sha256(
f.read_bytes()
).hexdigest()
for f in sorted(target.rglob("*"))
if f.is_file()
}
tasks.append({**row, "file_hashes": hashes})
local = (sources or {}).get(split)
manifest = {
"source_dataset": repo,
"source_revision": None if local else revision,
"expected_hub_revision": revision,
"local_source": str(local) if local else None,
"task_count": len(tasks),
"tasks": tasks,
}
write_json(destination / f"{split}_manifest.json", manifest)
manifests[split] = manifest
train = {r["source"] for r in manifests["train"]["tasks"]}
test = {r["source"] for r in manifests["test"]["tasks"]}
notebooks = [
{r.get("notebook") for r in manifests[s]["tasks"]} - {None}
for s in ("train", "test")
]
if train & test or notebooks[0] & notebooks[1]:
raise ValueError("Train/test overlap")
write_json(
destination / "ready.json",
{
"train": len(train),
"test": len(test),
"grader_sha256": GRADER_SHA256,
"train_test_overlap": 0,
},
)
return manifests
if __name__ == "__main__":
p = argparse.ArgumentParser(description=__doc__)
p.add_argument("--output", type=Path, default=ROOT / "prepared")
p.add_argument("--train-source", type=Path)
p.add_argument("--test-source", type=Path)
a = p.parse_args()
result = prepare(
a.output,
{s: v for s, v in (("train", a.train_source), ("test", a.test_source)) if v},
)
print(json.dumps({s: m["task_count"] for s, m in result.items()}))
|