early-experience-data

What problem does it solve?

This Skill addresses the challenge of generating SFT-ready training data for early-experience paradigms in agent environments, specifically for Implicit World Modeling (IWM) and Self-Reflection (SR) methods.

Core Features & Use Cases

• Data Generation: Automates the creation of expert, IWM, and reflection data for agent environments.
• Method Mapping: Guides the mapping of methods like IWM and SR to specific environments.
• Alternative Action Sampling: Provides strategies for sampling alternative actions in environments with enumerable or open action spaces.
• Reflection Generation: Facilitates the generation of self-reflection content for training policies.
• Use Case: Utilize this Skill to generate training data for an agent navigating a virtual world, such as predicting the next state or generating reasoning for expert actions.

Quick Start

Use the early-experience-data skill to generate expert and reflection data for a virtual environment by pointing the agent at skill/SKILL.md.