What it is
SFT dataset for reproducing early-experience agent training across 8 text-based environments with expert, reflection, and world-model traces. ~820k ShareGPT-format text trajectories across 8 agent environments and 3 training paradigms: expert imitation, self-reflection, and implicit world-model-to-expert supervision.
Key results
- 8 environments; WebArena has only iwm and reflection
- ~821k JSONL records in text-only ShareGPT format
- MIT-licensed; English; usable via Hugging Face datasets
Useful for supervised fine-tuning and reproducing LLM agent training with imitation, reflection, and world-model traces.