trainer

Trainer Skill

Purpose

Trainer Skill is the Codex-facing entry point and the canonical Python implementation for LoopAI model training.

Use this skill for:

• Running sft + llamafactory or grpo + verl; do not mix backend/stage pairs
• Validating SFT JSON/JSONL or GRPO Parquet data and reward configuration
• Generating and approving backend-specific training YAML
• Running or reconnecting to the persistent Trainer worker
• Reading local metrics, selecting checkpoints, and exporting Verl FSDP actors
• Returning structured Trainer errors without relying on SwanLab or Trainer MCP

Do not use this skill for Judger, Analyzer, data crawling, or broad project refactors.

Python Implementation

loopai/skills/Trainer/
├── __init__.py        # prepare() / run_prepared() / run() / result helpers
├── trainer_agent.py   # LangGraph subgraph used by Starter and the skill runner
├── nodes/             # data validation, config generation, training execution
├── rewards/           # stable LoopAI routing to Verl reward implementations
├── utils/             # persistent worker, Verl/SFT launchers, events, parsers
├── templates/         # bundled training templates
├── results.py         # parses metrics and selects the best checkpoint
├── runner.py          # skill entry that runs the Trainer subgraph
├── runtime_config.py  # resolves kwargs/env/state/starter.yaml
└── worker_entry.py    # independent process that owns training and finalization

The root skill description lives at:

skills/Trainer/SKILL.md

Mandatory YAML Approval

For every user-initiated training round, use this two-stage workflow:

1. Call prepare() to validate the data and generate the final training YAML. This stage must not start training.
2. Read result["trainer"]["trainer_result"]["data"] and show the user:
   • config_path
   • the complete config_yaml in a YAML code block
   • the selected train_framework and train_stage
   • for SFT: dataset/model paths, learning rate, epochs, batch size, LoRA fields, devices, output directory, and save_total_limit
   • for GRPO: train/validation Parquet paths, model, reward mode/preset or custom function, rollout backend, GPUs, batch/token limits, save/test frequency, checkpoint directory, selection metric, and checkpoint retention
3. Stop and wait for explicit user approval. Do not treat a previous round's approval as approval for a new round.
4. If the user requests edits, update the generated YAML, call inspect_prepared_config(), show the complete updated YAML, and wait for approval again.
5. Only after approval, call run_prepared() with the displayed config_path, its displayed config_sha256, and the trainer_version_id returned by prepare().
6. Keep run_prepared() in the foreground until training reaches completed, failed, or cancelled.