TanitAD Alpamayo-2-Super augmentation of PhysicalAI-AV
4,800 clips (~26.7 h of driving), 23,999 inference rows, 5 tasks per clip from
nvidia/Alpamayo2-Super run over a stratified selection of the NVIDIA
PhysicalAI-Autonomous-Vehicles dataset. NO raw sensor data is included: every row
links to the original clip by clip_id + t0_us (loader-default 5,100,000), so the
outputs can be joined back to camera/egomotion/label streams at any time.
Tasks (per clip, all outputs logged completely)
trajectory (sampled trajectory sets + minADE vs ego future), meta_action,
auto_labeling, vqa (questions drawn randomly from the 506-question bank in
vqa_bank_500.json), grounding_via_vqa. Raw model output preserved verbatim under
raw; provenance fields: model_id (nvidia/Alpamayo2-Super), seed,
_quantisation (4-bit bnb load), _contamination note, peak_gib.
Selection & coverage
Road-class stratification over the corpus (ego-dynamics-derived classes; no map
data exists in PhysicalAI): urban 1,884 · intersection-rich 1,241 · highway 384 ·
unstructured 83 · remainder mixed. Selection manifest: selection_manifest.json
(4,800 entries); per-clip road classes: aug_road_class.json in the TanitAD repo.
Files
records.parquet (all rows), vqa_bank_500.json, selection_manifest.json.
Known holes
One task row of 24,000 missing (23,999) after two MooseFS I/O incidents during production; per-(clip,task) completeness is auditable from the parquet. Torn-line tolerant readers recommended.
- Downloads last month
- 12