CoT chunk-K train set: all-step interleaved imagined reasoning; re-grounds on the true frame every K=3 steps. layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline. images are base64-encoded JPEG frames stored inline in each JSONL row.

Pairs with the matching pusht checkpoint(s) under the companion model org; CoT and non-CoT variants… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pusht_cot_chunk_k3_train.