Neptune-T/alphabot2-ovmm-sft

AlphaBot2 OVMM SFT data

Exact training inputs for Neptune-T/alphabot2-ovmm-sft3000.

Directory	Episodes	Frames / eligible anchors
alphabot2_ovmm_v2_train	2099	1756966
alphabot2_ovmm_v2_val	102	88577
nav_corrections/segments	32 NAV segments	9353

Original train and validation directories are separate LeRobot v3 datasets (10 Hz, three 240x320 RGB video streams, state16, Cartesian action18). Each directory retains data parquet, videos, metadata/statistics, frame_meta.npz, ovmm_episode_index.json and policy_io.json byte-for-byte. Their internal LeRobot splits both say train; use the directory and original episode index to distinguish the scientific training and validation split. Never merge validation episodes into training. These are SFT splits drawn from the physical Habitat TRAIN population, not the Habitat VAL benchmark.

from huggingface_hub import snapshot_download
root = snapshot_download("Neptune-T/alphabot2-ovmm-sft", repo_type="dataset", local_dir="ovmm-data")

The final model also used the 32 qualified NAV teacher-continuation segments at 20% sampling mass. Their first15 frames are history context, not eligible anchors; tail action masks must be preserved. Segment NPZ/JSON files and all six qualification evidence files are included unchanged. The original manifest preserves provenance. The portable manifest changes only paths. After download run python ovmm-data/prepare_local_manifest.py to create an absolute-path local manifest; use that path and its newly computed hash in your relocated training config. The original training manifest hash remains in model training provenance and must not be represented as the relocated hash. The correction loader is included in the model's training/code/.

No source episodes or statistics were regenerated for this publication. This is converted SFT data plus the correction supplement, not all original raw expert recordings or the Habitat/HSSD asset bundle. Upstream asset and dataset terms still apply; this copy adds no new license for upstream assets. MANIFEST.json records every payload file's size and SHA256.