tools/fetch_run6.py

"""One-shot script to pull Run 6 artifacts from HF Hub.

Written as a file (not python -c) so shells don't choke on the newlines.

Run 6 focus: break Run 5's degenerate R2-only policy via forced-outcome
task variants + unlikeliness reward shaping (He et al. 2506.02355) +
R-level balance bonus + curriculum with 50%→70% forced variants.

After training completes on the Space, run:

    python tools/fetch_run6.py

to pull every artifact locally for post-mortem. The destination folder is
`training_runs/run_6_forced_variants/` (gitignored).

Key files to inspect after fetch:
    grpo/training_log.json          — per-episode predicted vs actual R-level
    grpo/metrics.json               — mean reward, catastrophe count
    eval/results.json               — scripted vs sft_only vs grpo_trained
    grpo/_trainer/trainer_state.json — TRL's internal metrics, look at
                                      frac_reward_zero_std (target <40%)

Theory predictions (from config.yaml notes):
    * frac_reward_zero_std drops from 70% → <40%
    * confusion matrix has non-zero entries in all 5 R rows
    * mean reward 0.60-0.75 (lower than Run 5's 0.664)
    * eval accuracy 75-85% across R-levels (vs 100% R2-only)
    * task_log_cleanup solved (was unsolved in Run 5)
"""
from __future__ import annotations

import os
import shutil
import subprocess
from huggingface_hub import snapshot_download


TARGET_DIR = "training_runs/run_6_forced_variants"


def main() -> None:
    if os.path.exists(TARGET_DIR):
        shutil.rmtree(TARGET_DIR)
    token = subprocess.check_output(["hf", "auth", "token"], text=True).strip()
    path = snapshot_download(
        repo_id="chane335/permanence-artifacts",
        repo_type="dataset",
        local_dir=TARGET_DIR,
        token=token,
    )
    total = 0
    for root, _dirs, files in os.walk(path):
        for f in files:
            rel = os.path.relpath(os.path.join(root, f), path)
            if ".cache" in rel:
                continue
            size = os.path.getsize(os.path.join(root, f))
            total += size
            print(f"  {size:>12,} bytes  {rel}")
    print(f"TOTAL: {total/1e6:.1f} MB")
    print(f"\nNext: python -c \"import json; "
          f"print(json.load(open('{TARGET_DIR}/grpo/metrics.json')))\"")


if __name__ == "__main__":
    main()