code/Abforge_Training/scripts/backup_scratch_archive.py

#!/usr/bin/env python
"""Back up the lightweight ABForge scratch artifacts to a PRIVATE HF dataset repo.

Mirrors scripts/upload_abforge_models.py conventions:
  - token from HF_TOKEN env (source secrets/hf_slowguess.env first), NOT global login
  - verifies the token belongs to SlowGuess before doing anything
  - idempotent: upload_folder skips files already on the Hub

Covers: results (infer/*.jsonl), analysis, server-only scripts, run meta, logs.tar.gz,
and a curated data_local set. See docs/scratch_backup_manifest.md for the full rationale.

Usage:
  source $WORK_ROOT/secrets/hf_slowguess.env
  python scripts/backup_scratch_archive.py            # dry-run: print plan only
  python scripts/backup_scratch_archive.py --upload   # actually create repo + upload
"""
import argparse, fnmatch, os, sys
from huggingface_hub import HfApi

WORK = "/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training"
CODE = "/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training"
ORG = "SlowGuess"
REPO_ID = f"{ORG}/abforge-scratch-archive"

# server-only scripts that are NOT in git
SCRIPT_FILES = [
    f"{WORK}/run_inference_local.py",
    f"{WORK}/run_inference_local_tm_only.py",
    f"{WORK}/run_inference_bailian.py",
    f"{WORK}/run_inference_claude_cli.py",
    f"{WORK}/run_eval_v18_bench44.sh",
    f"{WORK}/run_eval_v18_retry.sh",
    f"{WORK}/env_misha.sh",
    f"{WORK}/scripts/case_diff_v17_v18.py",
    f"{WORK}/scripts/compare_v17_v18_bench44.py",
]

# curated local-only data (NOT on abforge-data, not trivially regenerable)
DATA_LOCAL = [
    f"{WORK}/data/rl_task1_25479_filt2_6.jsonl",
    f"{WORK}/data/bench_1000_rubric_v2_final_tagfixed.jsonl",
    f"{WORK}/data/bench_200_rubric_v2_final_tagfixed.jsonl",
    f"{WORK}/data/bench_44_rubric_v2.jsonl",
    f"{WORK}/data/bench_50_rubric_v2.jsonl",
    f"{WORK}/data/bench_150_remaining_rubric_v2.jsonl",
    f"{WORK}/data/bench_3_rubric_v2.jsonl",
    f"{WORK}/data/bench_data_4_subset50_simple.jsonl",
]

# (local_path, path_in_repo, allow_patterns)  — folders
FOLDER_UPLOADS = [
    (f"{WORK}/infer",          "results",  ["*.jsonl"]),                 # 144 MB of results, skips *_merged_hf
    (f"{WORK}/analysis",       "analysis", None),                        # ignores below
    (f"{WORK}/hydra_outputs",  "meta/hydra_outputs", None),
    (f"{WORK}/wandb",          "meta/wandb", ["*.json", "*.yaml", "*.csv", "*.txt", "*.log"]),
]
FOLDER_IGNORE = ["__pycache__/*", "*.pyc"]

# files staged elsewhere by the slurm wrapper (created before upload)
EXTRA_FILES = [
    (f"{WORK}/hf_upload/logs.tar.gz", "logs/logs.tar.gz"),               # wrapper tars logs/ here
    (f"{CODE}/docs/scratch_backup_manifest.md", "meta/scratch_backup_manifest.md"),
]


def plan():
    print(f"target repo (private dataset): {REPO_ID}\n")
    print("== folders ==")
    for src, dst, allow in FOLDER_UPLOADS:
        ok = "OK " if os.path.isdir(src) else "MISS"
        print(f"  [{ok}] {src}  ->  {dst}/   allow={allow}")
    print("== scripts -> scripts/ ==")
    for f in SCRIPT_FILES:
        print(f"  [{'OK ' if os.path.isfile(f) else 'MISS'}] {f}")
    print("== data_local -> data_local/ ==")
    for f in DATA_LOCAL:
        print(f"  [{'OK ' if os.path.isfile(f) else 'MISS'}] {f}")
    print("== extra files ==")
    for src, dst in EXTRA_FILES:
        print(f"  [{'OK ' if os.path.isfile(src) else 'MISS'}] {src}  ->  {dst}")


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--upload", action="store_true", help="actually create repo + upload (default: dry-run)")
    args = ap.parse_args()

    plan()
    if not args.upload:
        print("\n[dry-run] re-run with --upload to create the private repo and push.")
        return

    tok = os.environ.get("HF_TOKEN")
    if not tok:
        sys.exit("HF_TOKEN not set — source secrets/hf_slowguess.env first")
    api = HfApi(token=tok)
    who = api.whoami()["name"]
    if who != ORG:
        sys.exit(f"HF_TOKEN belongs to {who}, not {ORG} — aborting")

    api.create_repo(REPO_ID, repo_type="dataset", private=True, exist_ok=True)

    for src, dst, allow in FOLDER_UPLOADS:
        if not os.path.isdir(src):
            print(f"SKIP folder (missing): {src}"); continue
        print(f"\n=== folder {src} -> {dst}/ ===", flush=True)
        api.upload_folder(repo_id=REPO_ID, repo_type="dataset", folder_path=src,
                          path_in_repo=dst, allow_patterns=allow,
                          ignore_patterns=FOLDER_IGNORE, commit_message=f"archive {dst}")

    for group, dst in ((SCRIPT_FILES, "scripts"), (DATA_LOCAL, "data_local")):
        for f in group:
            if not os.path.isfile(f):
                print(f"SKIP file (missing): {f}"); continue
            print(f"upload {f} -> {dst}/{os.path.basename(f)}", flush=True)
            api.upload_file(repo_id=REPO_ID, repo_type="dataset", path_or_fileobj=f,
                            path_in_repo=f"{dst}/{os.path.basename(f)}",
                            commit_message=f"archive {dst}/{os.path.basename(f)}")

    for src, dst in EXTRA_FILES:
        if not os.path.isfile(src):
            print(f"SKIP extra (missing): {src}"); continue
        print(f"upload {src} -> {dst}", flush=True)
        api.upload_file(repo_id=REPO_ID, repo_type="dataset", path_or_fileobj=src,
                        path_in_repo=dst, commit_message=f"archive {dst}")

    print(f"\nDONE https://huggingface.co/datasets/{REPO_ID}", flush=True)


if __name__ == "__main__":
    main()