9T100E-autolabel

This dataset contains all 900 non-visual-OOD demonstrations (100 episodes per task) of the nine original v1 MomentMem benchmark tasks. No _v2 task is included. Videos, actions, states, task text and original LeRobot metadata are retained. The original privileged keyframe intervals are replaced by autoregressive VLM predictions.

Layout and training

lerobot_format/<task_name>/{data,meta,videos}

Download this repository as a dataset root. Point HF_LEROBOT_HOME at its lerobot_format subdirectory and select the nine task names below in the memory-based-VLA OpenPI training configuration. Compute/validate normalization assets for the selected data configuration before training; this repository does not silently substitute a pretrained normalization asset.

Each meta/keyframes.json has the existing training schema:

{"episodes": [{"episode_index": 0, "mem_boundary_pairs": [[24, 24], [132, 132]]}]}

The numbers above are illustrative. Every actual interval is a singleton [t,t], where t is a zero-based frame index in the original 20 Hz video, not an index in the 5 Hz sampled timeline. Empty lists are retained when the VLM selects no frames. History remains strictly causal: a selected frame is available to the VLA only at later timesteps. The standard VLA interval sampler can consume these labels unchanged.

Label-generation protocol

• Model: jellerode/MomentMem-VLM, step 10000, fixed 5 Hz, matching HF revision 79ff21bb998330c92123d4df68c516335b13f9c1.
• Source checkpoint basename: qwen_fixed5hz_fresh10k_step10000_20260808_221902.
• All 900 episodes, including both the original train and validation partitions.
• Overview RGB only, 224x224; fixed original indices 0,4,8,...; recent buffer 10; maximum long-term prompt history 48; maximum prompt length 16384.
• Yes if the two-token normalized P(Yes) > 0.5; otherwise No.
• Entirely autoregressive memory built from the model's own previous decisions.
• No ground-truth keyframe annotations, memory instructions, temporal augmentation, label smoothing, or post-hoc interval-based correction are used for labeling.
• The model was trained on the source benchmark's training split. These are in-domain automatic labels, not a zero-shot generalization evaluation.

Per-task provenance is stored in meta/vlm_keyframe_generation.json. provenance/ retains the full sampled-timeline binary decisions, generation configuration and code hashes. CHECKSUMS.sha256 covers all published payload files. Videos and data are uploaded as ordinary file contents and do not depend on filesystem symlinks or access to the original server.

Tasks

AD_scroll_discovery, AO_hidden_pattern, ASD_seal_checking, ASO_placement_imitation, AST_mine_sweeper, AT_drawer_match, SD_explore_cube, SO_spring_compression, ST_flash_beacon.

The original recordings use 20 Hz, 224x224 overview and wrist videos, and RT rendering. The wrist videos are preserved for downstream VLA use even though label generation uses only the overview camera.