rl-agentic-job-cleanup

⚠ Do not add comments to YAMLs. Report your recommendations directly to the supervisor.

rl-agentic-job-cleanup

After an RL job terminates, publish laion/<job_name>-<step>-<size> (weights at repo root), trace dataset penfever/<job_name>, and a Supabase models row (training_type=RL).

Rules

• hf upload, NEVER hf upload-large-folder (deprecated stub; deadlocks on HF LFS 429s). Wrap long uploads in tmux, not nohup.
• --private is a no-value flag — do NOT pass --private false. Default is PUBLIC; omit it.
• Run trace upload + parse_skyrl_metrics.py from the otagent conda env (the RL venv lacks google.cloud.storage + matplotlib).
• On Leonardo, login-node hf upload is SIGKILLed at ~100s — use the sbatch+tunnel upload pattern.

0. Cancel pending retries

squeue -u $USER --format='%.18i %.80j %.8T' | grep <job_name>
scancel <retry_job_ids>

1. Select the checkpoint — trailing-5 reward EMA

# NOTE: there is an empty exports/ at the base level — ignore it. Real HF-exportable ckpts are nested:
ls -lt $EXPERIMENTS_DIR/<job_name>/<job_name>/exports/ | head -10

Use the EMA of reward/avg_raw_reward over a trailing-5 window.

Rules:

• EMA across ALL chronological steps, regardless of chain restarts. Collect every .out and sort by trainer/global_step; do not compute it per-chain link.
• Standard 5-period EMA: α = 2/(5+1) = 1/3; EMA_n = α·reward_n + (1−α)·EMA_{n−1}, EMA_1 = reward_1.
• Never select the first saved checkpoint (global_step_5 with hf_save_interval: 5) — EMA not warmed up. Start from the second-saved step (typically 10).
• Among saved, aligned checkpoints (multiples of hf_save_interval, excluding the first), upload the highest EMA. If scancelled before a save-aligned max-step, cap at the latest saved multiple.

import json, glob, re
rewards = {}  # step -> avg_raw_reward
for fn in glob.glob(f"{EXP_DIR}/logs/*.out"):
    for line in open(fn):
        m = re.search(r'trainer/global_step":\s*(\d+).*avg_raw_reward":\s*([\d.eE+-]+)', line)
        if m:
            step, r = int(m.group(1)), float(m.group(2))
            rewards.setdefault(step, r)  # first-seen wins (chain links may overlap)
steps = sorted(rewards)
alpha = 1/3
ema = {}; prev = rewards[steps[0]]
for s in steps:
    prev = alpha * rewards[s] + (1 - alpha) * prev
    ema[s] = prev
SAVE_EVERY = 5  # match hf_save_interval
aligned_eligible = [s for s in steps if s % SAVE_EVERY == 0 and s >= 2 * SAVE_EVERY]
best = max(aligned_eligible, key=ema.get)
print(f"best EMA={ema[best]:.4f} at step={best} (reward at that step={rewards[best]:.4f})")

Upload the checkpoint at exports/global_step_<best>/.

2. Locate the W&B run (optional)

From the job logs / trainer_log.jsonl: https://wandb.ai/dogml/OpenThoughts-Agent/runs/<run_id>. (Jupiter has no W&B — omit.)

3. Flatten model files to the upload-dir root

UPLOAD_DIR=/e/scratch/jureap59/feuer1/upload_staging/<job_name>-<step>
mkdir -p $UPLOAD_DIR
cp $EXPORT_DIR/policy/* $UPLOAD_DIR/
ls $UPLOAD_DIR/   # safetensors, config.json, tokenizer files all at root

4. Copy the launch config for reproducibility

cp hpc/skyrl_yaml/<config_used>.yaml $UPLOAD_DIR/rl_config.yaml

5. Scan for secrets before upload

trufflehog filesystem $UPLOAD_DIR --no-update                                   # if installed
trufflehog filesystem $EXPERIMENTS_DIR/<job_name>/<job_name> --no-update         # logs/traces too
# fallback:
grep -rIE '(sk-[a-zA-Z0-9]{20,}|AKIA[0-9A-Z]{16}|ghp_[a-zA-Z0-9]{36}|hf_[a-zA-Z0-9]{34}|eyJ[a-zA-Z0-9._-]+)' $UPLOAD_DIR

Redact before proceeding (the wrapper emits a JSON finding record even when clean):

python -m scripts.harbor.secret_redaction "$UPLOAD_DIR"

6. Upload to HuggingFace — laion/<job_name>-<step>-<size>

Include the global step and base-model size suffix (-20-32B, -30-8B).

# tmux for long uploads. OMIT --private (no-value flag; default public).
hf upload laion/<job_name>-<step>-<size> $UPLOAD_DIR . --repo-type=model

The SkyRL trainer auto-pushes laion/<job_name> with weights under checkpoints/step_N/. Upload the manually flattened export to -<step>-<size> instead.

7. Register in the DB (--training-type RL) — with cross-user FK safety

Delete the trainer's auto-registered duplicate only if safe, then push the correct row. If any other-user row in sandbox_jobs, sandbox_trial_model_usage, or elsewhere FKs the auto-row, stop; do not delete or mutate the FK'd rows. Restrict all writes to rows you own.

other_users_fk = (c.table("sandbox_jobs").select("id,username,model_id")
    .eq("model_id", auto_row_id).neq("username", os.environ.get("USER","<you>")).execute())
if other_users_fk.data:
    print(f"SKIPPING auto-row delete — {len(other_users_fk.data)} other-user rows FK'd.")
else:
    c.table("models").delete().eq("name", "laion/<job_name>").execute()
    # optional, ONLY if pre-check passed: HfApi().delete_repo("laion/<job_name>", repo_type="model")

Then register the -<step>-<size> repo (--training-type RL is REQUIRED — the script defaults to SFT):

python scripts/database/manual_db_push.py \
  --hf-model-id laion/<job_name>-<step>-<size> \
  --base-model <base_model_hf> \
  --dataset-name <dataset_name> \        # comma-separated for multi-dataset → sets dataset_names
  --training-type RL                      # --wandb-run optional (defaults to now)

Verify --base-model is the exact HF repo trained from, not a default. Cross-check the job-name suffix, trainer.policy.model.path in the .out launch command, or notes/ot-agent/rl_experiments.md.

8. Upload RL traces → penfever/<job_name>

From the otagent env, always pass --skip_register for RL; register the model separately in step 7.

python -m scripts.harbor.make_and_upload_trace_dataset \
  --job_dir "$EXPERIMENTS_DIR/<job_name>/<job_name>" \
  --repo_id penfever/<job_name> --episodes last --skip_register

Never subsample or cap: upload the full trial set. The script reads the inner <job>/<job> trace_jobs/. For image-backed Jupiter runs, pass the same inner run root. The exporter detects artifact_store.img, refuses an unsafe mount while a writer lock is active, mounts it read-only after the link exits, and unmounts it when export finishes. Legacy bare trace_jobs/ runs keep the existing path.

make_and_upload_trace_dataset buffers the full dataset; chunk_size does not bound peak RAM. Large login-node uploads can OOM; do not respond by sampling.

Then add a "Training Traces" section to $UPLOAD_DIR/README.md (append if a model card exists, don't overwrite) linking penfever/<job_name>:

## Training Traces
Training-time Daytona/Harbor rollouts: **[penfever/<job_name>](https://huggingface.co/datasets/penfever/<job_name>)**
(the `last` episode of each trial — the rollouts the policy trained on after rollback/truncation).

9. Parse metrics and preserve training logs

python scripts/analysis/parse_skyrl_metrics.py \
  $EXPERIMENTS_DIR/<job_name>/logs $UPLOAD_DIR/training_logs \
  --trace_jobs_dir $EXPERIMENTS_DIR/<job_name>/<job_name>/trace_jobs
cp $EXPERIMENTS_DIR/<job_name>/<job_name>/trainer_log.jsonl $UPLOAD_DIR/training_logs/ 2>/dev/null
cp $EXPERIMENTS_DIR/<job_name>/logs/<job_name>_*.out $UPLOAD_DIR/training_logs/
hf upload laion/<job_name>-<step>-<size> $UPLOAD_DIR . --repo-type=model   # additive

Produces metrics.csv, vllm_metrics.csv, trial_stats.csv, report.md, reward_plot.png. The metrics reader also detects a missing <run>/trace_jobs beside artifact_store.img and mounts the inactive image read-only for trial statistics. WARNING: never use huggingface_hub.upload_folder() without delete_patterns=[] — it deletes files absent locally and clobbers the weights. hf upload is additive (safe).

10. Verify the published model and write the completion record

The model is complete only when its remote repo has weights, README.md with Training Traces, and redacted training_logs/. Record every artifact as present, absent, or not applicable in the cleanup handoff; an absent required artifact is not a completed cleanup.

hf api repo-info laion/<job_name>-<step>-<size> --repo-type model --expand siblings \
 | python -c 'import json,sys; files=[x["rfilename"] for x in json.load(sys.stdin)["siblings"]]; print({"weights": any(x.endswith(".safetensors") for x in files), "readme": "README.md" in files, "training_logs": any(x.startswith("training_logs/") for x in files)})'

Fetch README.md and verify it contains ## Training Traces and the exact penfever/<job_name> URL.

11. Clean up the experiments dir

After all prior steps succeed, rm -rf the local job dir. Detach a large GPFS removal with nohup or tmux; do not du or find it first.