early-experience-data - Generate early-experience SFT data
Generates, collects, inspects, and prepares early-experience training data for agent environments as expert, IWM, and reflection SFT JSONL files.
Tags
Updated: 2026-10-01Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Generate expert trajectories
- Generate IWM training data
- Generate reflection training data
- Run environment smoke tests
- Inspect generated samples
- Prepare SFT-ready JSONL
Inputs
- Agent environment repository
- METHOD.md
- method_recap.md
- pitfalls.md
- Environment NOTES.md
- Generator model choice
- User batch approval
- User filter approval
Outputs
- Expert SFT JSONL files
- IWM SFT JSONL files
- Reflection SFT JSONL files
- Environment pipeline records
Requirements
- LLM API access
- A supported generator model
- A runnable agent environment
- Access to the native evaluator