novastar112/pusht_96_int1_visual_nomarker_allstep_thinking_trickiness_cot

PushT int1 Visual Nomarker All-Step Thinking Trickiness COT

This dataset is derived from successful PushT visual-nomarker trajectories in novastar112/pusht_96_int1_visual_nomarker. Each row contains one full successful trajectory from the first move through the final stop action.

Main files:

• —training/pusht_allstep_thinking_cot.jsonl.gz: 500,000 train rows.
• —testing/pusht_allstep_thinking_cot.jsonl.gz: 200 test rows.

Message format:

• —Each user turn is the PushT prompt text plus one current image item with loss=false.
• —The user image item stores image_prev, image, and image_next; image_prev == image is validated for every step.
• —One selected max-coverage-jump non-stop move step contains four imagined one-step branch images with loss=true.
• —All other non-stop assistant turns use a fixed trivial thinking template.
• —The final assistant turn uses a fixed stop thinking template and outputs ('stop', 'stop').

Key-step policy:

• —Replay the source trajectory and select the non-stop move step with the largest one-step coverage change.
• —Use a 48 px PushT-world perturbation radius.
• —Prefer motion-frame perturbations around the planned target: planned, shortertarget, farthertarget, leftpath, rightpath.
• —Fall back to goal-frame perturbations only if the motion-frame trial set is not contrastive.
• —Keep planned target plus the three strongest contrastive perturbations, for four branch images total.

Target Hub repo: https://huggingface.co/datasets/novastar112/pusht96int1visualnomarkerallstepthinkingtrickinesscot