agentic-in Training
Use this built-in skill when the user is planning or executing a fine-tuning, post-training, or model adaptation workflow.
Core rules
- Start from the training objective and data posture, not the framework choice.
- Confirm hardware, checkpoint, tokenizer, and evaluation compatibility up front.
- Keep launch configs, dataset versions, and output checkpoints traceable.
- Treat rollback and validation as part of the training plan.
Default workflow
- Define the target behavior change and evaluation criteria.
- Inspect dataset quality, formatting, and licensing constraints.
- Choose the training path that matches the budget and hardware.
- Validate checkpoints against the baseline before promoting them.
Guardrails
- Do not launch expensive runs without a measurable success condition.
- Do not assume dataset quality from format alone.
- Do not promote a checkpoint that has not been evaluated against the baseline.