Rl Agentic Job Cleanup

Clean up an agentic RL run

Complete every independent preservation step even when another step is blocked. Record which steps completed, did not apply, or could not proceed.

Authority

Publishing the model and companion trace dataset is part of this workflow when the campaign record names their destinations. Routine ingress and egress within the selected CoreWeave job scope do not require confirmation for each action. Registration is optional and requires campaign authorization. Deletion and storage reclamation require separate explicit authority and happen only after destination checks pass. Large or cross-region transfers remain subject to repository policy. Never invent a namespace, registry target, base model, or campaign policy.

Inputs

Read the terminal job chain, resolved configuration, checkpoint interval, artifact locations, base model, publication destinations, and registration policy from current records. Cancel only exact pending retries that belong to this run before reading mutable checkpoint directories.

Workflow

1. Capture the chain. Sync every attempt's logs and durable artifacts. Preserve the tracker URL when one exists.

2. Select a checkpoint. Use infra/rl_cleanup/parse_skyrl_metrics.py across the chronological chain to build the metric surface, passing --trace_jobs_dir when traces are stored elsewhere. The parser detects training-log serialization independently of the agentic harness. The tool does not select an agentic checkpoint. Apply the campaign's declared selection rule only to complete saved checkpoints. If no rule is declared, report the valid candidates and ask for a choice. When the run ended at a known behavioral break, include the last checkpoint whose generating rollouts predate that boundary.

3. Export or consolidate. Determine the checkpoint format before converting. Follow .agents/ops/checkpoint-consolidation.md and verify that the result contains complete safetensors, configuration, and tokenizer metadata.

4. Publish the model. Assemble the metric surface, trainer log, and relevant per-step logs under <staging>/training_logs/ before upload. Add a ## Training Traces section to the model card with the companion Hugging Face dataset URL; this is a model-publication requirement even when trace publication finishes in a later pass. Redact the complete staging tree before upload:

   python -m infra.rl_cleanup.secret_redaction "$STAGING_ROOT"
   hf upload <namespace>/<model> "$STAGING_ROOT" --repo-type model
   hf download <namespace>/<model> --repo-type model --local-dir "$VERIFY_ROOT"
   python -m infra.rl_cleanup.secret_redaction "$VERIFY_ROOT" --check
   python -m infra.rl_cleanup.publication_checks "$VERIFY_ROOT"
   

   Keep staging until the fresh download passes both checks. The first check fails if credential-shaped text remains; the second fails unless training_logs/ is nonempty and the card links a Hugging Face dataset under ## Training Traces.

5. Publish traces. The monitor's trace_jobs/ directory is an evidence sample by default; it mirrors only the newest 500 traces across the fleet. Rebuild the dataset from the full durable object-store tree instead:

   python infra/sync_rl_logs.py /benjaminfeuer/<job> --cluster <cluster> --dest "$TRACE_ROOT" \
     --no-ray --no-finelog --trace-jobs --trace-jobs-no-gzip
   mkdir -p "$TRACE_ROOT/trace_jobs"
   tar -xf "$TRACE_ROOT/<job>_trace_jobs.tar" -C "$TRACE_ROOT/trace_jobs"
   python -m infra.rl_cleanup.make_and_upload_trace_dataset \
     --job_dir "$TRACE_ROOT/trace_jobs" --repo_id <namespace>/<dataset> --episodes last --filter none \
     --skip_register --single_commit
   

   --trace-jobs is the source of record for this rebuild. It writes a tar plus a sync manifest; inspect objects_skipped before export. A full monitor mirror requires --trace-sync-limit 0, but remains evidence-only when its size guard skips objects. Inspect trace_export_manifest.json: result_coverage is dataset rows divided by source result.json files and defaults to a 95% gate. The staged tree has deterministic shard names and replaces stale train-*.parquet shards on rerun. --single_commit stages then replaces the complete remote shard set. Use --stage-only "$STAGING_ROOT" to retain that staged tree locally for inspection or offline tests; it does not write to Hugging Face. Verify the remote shard list and manifest before reporting the trace destination to the model-publication step.

6. Register if authorized. Use the campaign's schema, explicit RL training type, exact base-model lineage, and only records the operator owns. Stop on foreign-key or ownership ambiguity.

7. Reclaim only if separately authorized. Retain staging and source artifacts until all remote verification succeeds. Avoid expensive recursive sizing before a large delete.

Safety rules

• A training checkpoint is not a publishable model until the export is verified.
• An automatically pushed intermediate with nested weights is not a substitute for the flat model.
• Do not let one failed upload suppress metrics, traces, or other preservation work.
• Do not use destructive upload semantics or remove remote files absent from local staging.
• Never infer model size, base lineage, or checkpoint eligibility from a run name.

Completion record

Report terminal state, source revision, selected checkpoint and rationale, export provenance, published model and trace destinations, registration result, preserved metrics, verification evidence, storage reclaimed under authority, and all remaining work. For each required model artifact, state present, absent, or not applicable: weights, tokenizer/configuration, training_logs/, and the model-card Training Traces link. A model is incomplete until every applicable artifact is present.