#!/usr/bin/env python3 """Remove the legacy `Rubric_v1_backup` field from the ABForge HF dataset, and (optionally) refresh the dataset card. Usage: # 1) make sure the cached HF token is SlowGuess's (write-enabled), e.g. # python -c "from huggingface_hub import login; login()" # paste token # # 2) dry run -- only reports which files contain the field (no download of # big files, no changes pushed): DRY_RUN=1 /path/to/conda/abforge/bin/python remove_rubric_v1_backup.py # # 3) apply -- downloads matching files, strips the field, re-uploads, and # updates the card, all in ONE commit: DRY_RUN=0 /path/to/conda/abforge/bin/python remove_rubric_v1_backup.py """ import os import json from pathlib import Path import requests from huggingface_hub import HfApi, hf_hub_download, CommitOperationAdd REPO = "SlowGuess/abforge-data" FIELD = "Rubric_v1_backup" CARD_LOCAL = "/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/abforge_dataset_card.md" DRY = os.environ.get("DRY_RUN", "1") != "0" token = Path.home().joinpath(".cache/huggingface/token").read_text().strip() api = HfApi(token=token) who = api.whoami()["name"] print(f"Authenticated as: {who}") if who != "SlowGuess": print("!! WARNING: not SlowGuess -- writes will fail or go to the wrong repo.") files = [f for f in api.list_repo_files(REPO, repo_type="dataset") if f.endswith(".jsonl")] def first_row(rel): """Fetch just the first line via an HTTP Range request (no full download).""" url = f"https://huggingface.co/datasets/{REPO}/resolve/main/{rel}" r = requests.get(url, headers={"Authorization": f"Bearer {token}", "Range": "bytes=0-5000000"}, timeout=120) line = r.text.split("\n", 1)[0] return json.loads(line) # --- detect affected files (cheap: first row only) --- affected = [] print("\n== Detection (first-row probe) ==") for rel in files: try: keys = list(first_row(rel).keys()) except Exception as e: print(f" {rel}: could not probe ({e})") continue has = FIELD in keys other_v1 = [k for k in keys if "v1" in k.lower() and k != FIELD] note = f" [other v1-ish keys: {other_v1}]" if other_v1 else "" print(f" {rel}: {FIELD}={'YES' if has else 'no'}{note}") if has: affected.append(rel) print(f"\nAffected files: {affected or '(none)'}") if not affected: print("Nothing to do for the data. (Card can still be updated below.)") if DRY: print("\nDRY RUN -- no files downloaded/rewritten, no commit. " "Re-run with DRY_RUN=0 to apply.") raise SystemExit(0) # --- apply: download, strip field, rewrite, collect ops --- ops = [] for rel in affected: print(f"\n-- processing {rel} --") local = hf_hub_download(REPO, rel, repo_type="dataset", token=token) out = local + ".clean" removed = 0 total = 0 with open(local, "r", encoding="utf-8") as fin, open(out, "w", encoding="utf-8") as fout: for line in fin: line = line.rstrip("\n") if not line: continue total += 1 d = json.loads(line) if FIELD in d: del d[FIELD] removed += 1 fout.write(json.dumps(d, ensure_ascii=False) + "\n") print(f" rows={total}, removed {FIELD} from {removed} rows -> {out}") ops.append(CommitOperationAdd(path_in_repo=rel, path_or_fileobj=out)) # refresh the card too (YAML header + no v1 mention) if os.path.exists(CARD_LOCAL): ops.append(CommitOperationAdd(path_in_repo="README.md", path_or_fileobj=CARD_LOCAL)) print(f"\nWill also update README.md from {CARD_LOCAL}") if ops: info = api.create_commit( repo_id=REPO, repo_type="dataset", operations=ops, commit_message="Remove legacy Rubric_v1_backup field; refresh dataset card", ) print("\nCOMMIT DONE:", getattr(info, "commit_url", info)) else: print("\nNo operations to commit.")