This is the LIBERO dataset (1693 episodes, 40 tasks) augmented with SARM dense_only subtask annotations. Each episode's high-level task instruction is decomposed into ordered atomic subtasks (from decompositions.json), and a vision-language model (Qwen/Qwen3-VL-8B-Instruct) localizes each subtask in time. Per-episode columns added to meta/episodes/*.parquet:

dense_subtask_names, dense_subtask_start_times… See the full description on the dataset page: https://huggingface.co/datasets/k1000dai/libero_subtask_sarm.