VibeHash's picture
Release laya-ledgar, run ce-seed0
779ff0b verified
Raw History Blame Contribute Delete
5.4 kB
"""Download a task's source at a pinned revision, check it, and write the files Laya trains and is scored on.
The source is pinned by commit and by SHA-256, so a rerun reads the same bytes. The test file this script writes is
locked by its own SHA-256 as well. If anything about how a test question is written changes (the option texts, the
instruction, the order), the hash changes and the script refuses to go on, because scores made before and after the
change would no longer compare.
Run it with `python -m layatasks.data <task>`.
"""
from __future__ import annotations
import argparse
import json
from collections import Counter
from pathlib import Path
from typing import Any, Dict, Iterable, List, Sequence, Tuple
from layatasks import task as tasks
from layatasks.common import data_dir, normalise, sha256_file, write_json, write_jsonl
def overlap(train_texts: Iterable[str], test_texts: Iterable[str]) -> int:
"""How many test texts also appear in training, ignoring case and spacing."""
seen = {normalise(t) for t in train_texts}
return sum(1 for t in test_texts if normalise(t) in seen)
def download(task) -> Dict[str, Path]:
"""Each source file at the pinned revision, checked against its SHA-256."""
from huggingface_hub import hf_hub_download
paths = {}
for split, (filename, expected) in task.files.items():
path = Path(hf_hub_download(task.repo, filename, repo_type="dataset", revision=task.revision))
got = sha256_file(path)
if got != expected:
raise RuntimeError("%s: SHA-256 is %s, expected %s" % (filename, got, expected))
paths[split] = path
return paths
def read_parquet(path: Path, label_column: str) -> Tuple[List[Dict[str, Any]], List[str]]:
"""`(records, label names)`. The names come from the file's own metadata, in label-id order."""
import pyarrow.parquet as pq
table = pq.read_table(path)
features = json.loads(table.schema.metadata[b"huggingface"])["info"]["features"]
return table.to_pylist(), list(features[label_column]["names"])
def check_test_lock(task, path: Path) -> str:
"""Refuse a test file whose hash moved. Returns the hash."""
got = sha256_file(path)
if task.test_sha256 is not None and got != task.test_sha256:
raise RuntimeError("%s is %s, locked at %s. The test changed; earlier scores no longer compare."
% (path.name, got, task.test_sha256))
return got
def label_counts(records: Sequence[Dict[str, Any]], names: Sequence[str], column: str) -> Dict[str, int]:
counts = Counter(r[column] for r in records)
return {names[i]: counts.get(i, 0) for i in range(len(names))}
def build(task, out: Path) -> Dict[str, Any]:
out.mkdir(parents=True, exist_ok=True)
sources = download(task)
train_records, names = read_parquet(sources["train"], task.label_column)
test_records, test_names = read_parquet(sources["test"], task.label_column)
if names != test_names:
raise RuntimeError("train and test name their labels differently")
write_jsonl(out / "train.jsonl", task.rows(train_records, names))
write_jsonl(out / "test.jsonl", task.rows(test_records, names))
test_sha = check_test_lock(task, out / "test.jsonl")
write_json(out / "labels.json", {"names": names, "options": task.options(names)})
train_counts = label_counts(train_records, names, task.label_column)
test_counts = label_counts(test_records, names, task.label_column)
manifest = {
"task": task.name,
"source": {"repo": task.repo, "revision": task.revision, "license": task.license,
"attribution": task.attribution,
"files": {s: {"path": p, "sha256": h} for s, (p, h) in task.files.items()}},
"question": {"id": task.question_id, "instructions": task.instructions, "options": len(names)},
"train": {"rows": len(train_records), "per_label_min": min(train_counts.values()),
"per_label_max": max(train_counts.values()), "per_label": train_counts},
"test": {"rows": len(test_records), "sha256": test_sha, "locked": task.test_sha256 is not None,
"per_label_min": min(test_counts.values()), "per_label_max": max(test_counts.values()),
"per_label": test_counts,
"also_in_train": overlap((r[task.text_column] for r in train_records),
(r[task.text_column] for r in test_records))},
}
write_json(out / "manifest.json", manifest)
return manifest
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
parser.add_argument("task", choices=tasks.TASKS)
parser.add_argument("--out", type=Path, default=None, help="output folder (default: $LAYA_TASKS_HOME/<task>/data/v1)")
args = parser.parse_args()
task = tasks.load(args.task).TASK
manifest = build(task, args.out or data_dir(task.name))
short = {k: ({kk: vv for kk, vv in v.items() if kk != "per_label"} if isinstance(v, dict) else v)
for k, v in manifest.items()}
print(json.dumps(short, indent=2))
if not manifest["test"]["locked"]:
print("test.jsonl is not locked yet: set test_sha256 = %r in layatasks/tasks/%s.py before scoring any model"
% (manifest["test"]["sha256"], task.name))
if __name__ == "__main__":
main()