Datasets:
Download scripts/assemble_heldout.py from NagaYu/canary-eval: direct link, hf CLI and curl.
- Browser
- Download file 2.63 kB
-
https://huggingface.co/datasets/NagaYu/canary-eval/resolve/main/scripts/assemble_heldout.py
- Command line
-
hf download hf://datasets/NagaYu/canary-eval/scripts/assemble_heldout.py
-
curl -L -o assemble_heldout.py https://huggingface.co/datasets/NagaYu/canary-eval/resolve/main/scripts/assemble_heldout.py
2.63 kB
| #!/usr/bin/env python3 | |
| """Rebuild the held-out test split from the writer/reviewer records. | |
| A case is kept only if its contract is valid, the blind reviewer gave the same label as the writer, | |
| and the reviewer was confident. Run from the root of the Canary repository (it imports app.py only to | |
| validate contracts; no verdict is computed here): | |
| python hf/dataset/scripts/assemble_heldout.py hf/dataset/provenance/heldout_writer_reviewer.json /tmp/test.jsonl | |
| python hf/dataset/scripts/flatten.py /tmp/test.jsonl | |
| """ | |
| import collections | |
| import json | |
| import logging | |
| import re | |
| import sys | |
| JA = re.compile(r"[-ヿ㐀-鿿]") | |
| def main() -> None: | |
| sys.path.insert(0, ".") | |
| logging.disable(logging.CRITICAL) | |
| import app | |
| records = json.load(open(sys.argv[1], encoding="utf-8")) | |
| kept, dropped = [], collections.Counter() | |
| for rec in records: | |
| c, r = rec["writer"], rec["reviewer"] | |
| contract = dict(c["contract"]) | |
| contract.setdefault("type", c["contract_type"]) | |
| if r.get("label") is None: | |
| dropped["no review"] += 1 | |
| elif contract.get("type") != c["contract_type"]: | |
| dropped["type mismatch"] += 1 | |
| elif not app.validate_contracts([dict(contract, id="C", severity="critical")])["valid"]: | |
| dropped["invalid contract"] += 1 | |
| elif c["contract_type"] == "stable" and not isinstance(c["output"], list): | |
| dropped["stable without samples list"] += 1 | |
| elif r["label"] != c["label"]: | |
| dropped["writer and reviewer disagree"] += 1 | |
| elif not r.get("confident"): | |
| dropped["reviewer not confident"] += 1 | |
| else: | |
| text = json.dumps([c["output"], c["case"]], ensure_ascii=False) | |
| kept.append({ | |
| "contract_type": c["contract_type"], | |
| "contract": {k: v for k, v in contract.items() if k not in ("id", "severity")}, | |
| "output": c["output"], | |
| "case": c["case"] or {}, | |
| "label": c["label"], | |
| "lang": "ja" if JA.search(text) else "en", | |
| "category": c["category"], | |
| "difficulty": c["difficulty"], | |
| "rationale": c["rationale"], | |
| "source": f"held-out writer:{rec['group']}, blind reviewer agreed", | |
| }) | |
| with open(sys.argv[2], "w", encoding="utf-8") as fh: | |
| for i, row in enumerate(kept, 1): | |
| fh.write(json.dumps({"id": f"test-{i:04d}", **row}, ensure_ascii=False) + "\n") | |
| print("records", len(records), "kept", len(kept), "dropped", dict(dropped)) | |
| if __name__ == "__main__": | |
| main() | |