areno-run-training

name	areno-run-training
description	Run, configure, retry, and validate AReno SFT, DPO, GSPO, GRPO, PPO, and agentic training. Use for training commands, dataset or reward setup, smoke validation, real-step execution, checkpoint saving, or failed training retries. Do not use for serving-only tasks or framework implementation work.

Run AReno Training

Read repository AGENTS.md, CODEMAP.md, and current areno train --help before building a command.

For remote model or dataset references, explicitly pass --model-hub modelscope. Do not substitute a Hugging Face download when the ModelScope asset is missing; request a valid ModelScope ID or local path.

Select the path

Read references/algorithm-matrix.md. Inspect a bounded dataset sample with:

python .agents/skills/areno-run-training/scripts/inspect_dataset.py \
  --dataset-path <path-or-ref> --model-hub modelscope \
  [--loader examples/.../dataset_loader.py] --algo <algo>

Do not build or run the training command until this inspection returns "ok": true. If raw rollout rows lack prompt or messages, select a dataset loader and rerun the inspection with --loader; pass the same path to training as --dataset-loader-fn. For GSM8K-style question/answer rows, use examples/math/dataset_loader.py.

Use scripts/read_metrics.py to inspect event keys or selected scalar series. Do not parse stdout as the metric source.

Workflow

1. Record git rev-parse HEAD, environment facts from areno env --json and areno check, GPU state, checkpoint source, ModelScope dataset source, and resolved local paths.
2. Classify SFT, DPO, rollout RL, or agentic RL. Read references/data-contracts.md.
3. Inspect both the raw schema and the normalized schema. Treat a missing rollout prompt/messages as a required-loader error, not a warning.
4. Build the smallest command expressing the requested real workload. Preserve user-provided max_new_tokens and max_context_len, and include the verified loader with --dataset-loader-fn.
5. Use smoke or tune only when useful. Smoke is capacity evidence, not task completion.
6. Run the real job. Confirm the requested trainer step advances. For rollout, inspect one coherent sample and reward.
7. On failure, use references/failure-triage.md. Fix the first causal error.
8. If saving is requested, verify output and reload it through the intended adapter.

Capacity invariants

• batch_size * n_samples is total sample demand; max_running_prompts is concurrent active capacity.
• Rollout memory follows cache/context/concurrency. Train memory follows mini_bs, sequence length, activation and optimizer state.
• --drop-rollout-state changes lifecycle memory, not task semantics.
• Reduce concurrency or microbatch before semantic token limits.

Completion evidence

Report command, commit, model/dataset, topology, observed step and key metrics, plus save/reload evidence when required. Model load or smoke alone is not completion.