SmolVLA fine-tuned twice, once on Community Curated v0 as released and once on an edited version with the 1,365 episodes whose cameras show the hand and the object together on under half the labelled frames removed. Same budget, same eight task strings taken from the release itself. The edited version scored 5.1 points higher over 960 graded rollouts.
A same-size random cut scored 32.6%, so the gain is from which episodes were removed, not from removing episodes.
The chess move loses 449 episodes to the cut, more than any other task string in the release. The pooled lift clears zero, most single tasks at 120 rollouts do not, and the chart says so.
Measured on the whole release, 981 sub-datasets and 52,507 episodes and 229.5 hours, summed from every meta/info.json. Each of the 50,592 episodes carrying a canonical judgment gets one number, the share of its judge-labelled frames in which the gripper and the manipulated object are both in shot, and 1,365 episodes fall below half, 160 of them at zero, where the work is never in shot at all. This is not a re-run of the release's own keep and discard field, because 628 of those 1,365 carry a keep verdict. It lands disproportionately on one-camera recordings, which are 11.0% of the judged corpus and 51.4% of the removed part.
| Removed | episodes | mean on-camera |
|---|---|---|
| object and gripper never both in shot | 160 | 0% |
| in shot on under a quarter of the labelled frames | 127 | 16% |
| in shot on a quarter to half of them | 1,078 | 37% |
| Removed | 1,365 | 30% |
| Kept | 49,227 | 95% |
Measured on the release. The number on each card is that episode's on-camera share, the fraction of its judge-labelled frames in which the gripper and the object are both in shot, and every removed card below sits at zero. The last two cards are kept on purpose, including one the release's own judge marked review and logged five events against, because an imperfect episode the camera saw is something a policy can learn from and an empty frame is not.