NoahWeiss/large_home_dataset, measured from the public release on
18 September 2026.
Every row is read out of the release. Nothing here is estimated and nothing is rounded up.
| episodes declared in meta/info.json | 1,153 |
| camera features declared | observation.images.base and observation.images.wrist |
| episode parquet files present | 1,153 |
| base camera videos present | 1,153 |
| wrist camera videos present | 897 |
| episodes with no wrist video | 256, which is 22.2% |
| first missing episode indices | 19, 25, 31, 40, 41, 42, 43, 44, 45, 50 |
| large_home_dataset_v2, paths in common | 3,208 of 3,208 |
| large_home_dataset_v2, paths unique to either side | 0 |
| large_home_dataset_v2, files with a different byte size | 0 |
| farm_uf850_home_full, wrist videos present | 1,153 of 1,153 |
| farm_uf850_home_full, extra files against large_home_dataset | 256, every one a wrist video |
A policy that reads two camera streams gets one on 22.2% of these episodes. Depending on the loader that either drops those episodes without saying so or trains the wrist encoder on a repeated frame. The release that has every wrist video already exists, which is farm_uf850_home_full, so the fix is to point people at it. The copy named v2 is not a later version, it is the same 3,208 files at the same sizes, and it carries the same 256 gaps.
List the repo files and count videos/chunk-000/observation.images.wrist against videos/chunk-000/observation.images.base. 897 against 1,153. No credentials and no download.
This page is what the release tells you on its own. A full grading removes the part that scores worst, retrains on what is left, and reports whether the smaller corpus scores higher, with rollout counts and intervals. The one result measured on physical hardware is Stera-10M, two arms on a four arm SO-ARM101 bench, 21.9% against 36.9% task success across 1,270 rollouts. The Stera-10M report.