SmolVLA fine-tuned twice, once on community_dataset_v3 as released and once on an edited version with the 247 recordings whose task sentence was never written for them removed. Same budget, same eight tasks taken from the release itself. The edited version scored 5.3 points higher over 960 graded rollouts.
A same-size random cut scored 37.4%, so the gain is from which recordings were removed, not from removing recordings.
The tasks that gain most are the ones whose training sentence stopped being shared with unrelated behaviour once the cut ran. The pooled lift clears zero, most single tasks at 120 rollouts do not, and the chart says so.
Measured on all 1,755 recordings in the release. Each recording that carries a single task sentence gets one number, how many other contributors ship the byte-identical sentence, and the distribution has a hole in it. 1,274 recordings sit at zero, 255 sit at 164, and nothing at all sits between 5 and 163. Those 255 all carry the same sentence, Grasp a lego block and put it in the bin, so the cut takes the 247 of them whose own directory name does not independently name a lego block, brick or cube, and 196 of those 247 are still called so100_test or so101_test, and so101_test is the repo id the LeRobot record walkthrough puts in its own example. The cut leaves in the 302 recordings under five episodes that still carry their own sentence, and every failed grasp, because a bad demonstration filed under a sentence somebody wrote for it is still something a policy can read.
| Removed | recordings | episodes | hours |
|---|---|---|---|
| named exactly so100_test or so101_test | 98 | 1,387 | 8.2 |
| a suffixed variant of that same name | 98 | 1,078 | 5.9 |
| another name that never says lego | 48 | 1,418 | 7.6 |
| a name that says pens or bottle, not lego | 3 | 45 | 0.6 |
| Removed | 247 | 3,928 | 22.3 |
| Kept | 1,508 | 68,490 | 342.8 |
Measured on the release. Each card is one episode from a removed recording, cut at that episode's own from_timestamp inside the packed video, and the title is the sentence the episode actually carries. The last two cards are the contrast, ordinary recordings whose sentence nobody else in the release uses.