Trajectory data for the TCOD-v1 (temporal-curriculum on-policy distillation) ALFWorld experiments.
Used to train the SFT behavior-cloning baseline and as the teacher-prefix source for TCOD-b2f.
alfworld/teacher_rollout.jsonl
3,553
GiGPO-Qwen2.5-7B teacher pass@10 successful trajectories on ALFWorld train games. Each row: {game_file, target, actions} (bare actions). 124 rows have empty actions (hard games… See the full description on the dataset page:
https://huggingface.co/datasets/explcre/tcod-v1-alfworld-data.