| """Copy out what the evaluation reads from a committed run (the run stays). |
| |
| Usage: uv run python claude_run/NEW/scripts/02_extract.py <workspace> <run_id> |
| |
| Writes predictions/<run_id>.jsonl.gz with exactly the rows |
| `ablations/evaluate_step.py` hands to the evaluator (segment rows with |
| projection.include and no alternative_group; run_status rows) and a trimmed |
| copy of the manifest. |
| """ |
| import os |
| import gzip |
| import json |
| import sys |
| from pathlib import Path |
|
|
| ROOT = Path(os.environ.get("SEG_ROOT") or Path(__file__).resolve().parents[1]).resolve() |
|
|
|
|
| def main(): |
| ws, run = Path(sys.argv[1]), sys.argv[2] |
| rd = ws / "runs" / run |
| manifest = json.loads((rd / "manifest.json").read_text()) |
| out = ROOT / "predictions"; out.mkdir(exist_ok=True) |
| n = {"segment": 0, "run_status": 0, "excluded": 0} |
| with gzip.open(out / f"{run}.jsonl.gz", "wt") as fh, (rd / "projection.jsonl").open() as src: |
| for line in src: |
| r = json.loads(line); r = r.get("current", r) |
| if r["record_type"] == "run_status": |
| fh.write(json.dumps({"record_type": "run_status", "document_id": r["source"]["document_id"], "run_status": r["run_status"]}) + "\n") |
| n["run_status"] += 1; continue |
| if r["record_type"] != "segment": continue |
| if not r["projection"]["include"] or r["annotation"]["alternative_group"]: |
| n["excluded"] += 1; continue |
| fh.write(json.dumps({"record_type": "segment", "document_id": r["source"]["document_id"], |
| "text_sha256": r["source"]["text_sha256"], "ranges": r["source"]["ranges"]}) + "\n") |
| n["segment"] += 1 |
| keep = {k: manifest.get(k) for k in ("run_id", "stage", "parent_run_id", "additional_parents", "parameters", "status", |
| "change_counts", "record_count", "created_at", "code_sha256", "metadata", "manifest_sha256")} |
| (out / f"{run}.manifest.json").write_text(json.dumps(keep, indent=1, default=str)) |
| print(run, json.dumps(n)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
