Checkpoints¶
| Checkpoint | Backbone | Trained on | A-OKVQA | ScienceQA | VQAv2 yes/no | score head |
Params | Latency, L4 bf16 |
|---|---|---|---|---|---|---|---|---|
| thaitea/laya-vision, recommended (same weights as laya-vision-201m) | SmolVLM-256M cut to 20 of 30 language layers, options attend to each other | 19 Cauldron subsets + 4 rubric-scored sets + game frames | 59.8% | 82.4% | 71.4% | trained | 201M | 41 ms |
thaitea/laya-vision-smolvlm-256m-score (thaitea/laya-vision until 2026-09-24, revision d1fbdc0) |
SmolVLM-256M, options attend to each other | 19 Cauldron subsets + 4 rubric-scored sets | 60.0% | 82.8% | 72.4% | trained | 237M | ~41 ms |
Accuracies are calibrated answers on the official validation splits (VQAv2 yes/no is a re-split of the official
val set by image, so not comparable to published VQAv2 numbers). Latency is the median predict call on an L4,
preprocessing included; timed on the same GPU, the recommended checkpoint takes 0.83× the time of the previous one.
The recommended checkpoint¶
SmolVLM-256M-Instruct cut to its first 20 language-model layers, continued from the previous checkpoint and trained for 2 hours (28,414 steps at batch 64 on one H100) on the Cauldron and rubric sets plus game frames, which were 45% of training draws. The recipe came out of the autoresearch loop.
- Against the previous checkpoint:
- accuracy: 69.1% against 69.3% over 59,427 questions on 34 validation sets;
- within a point on 23 sets, ahead on 3 and behind on 8, mostly small ones (VQA-RAD, 62 questions, 80.6% against 88.7%).
- Calibration: ECE 0.041 pooled over all questions, against 0.064 for the previous checkpoint. By set: 0.092 on A-OKVQA, 0.028 on ScienceQA and 0.064 on VQAv2 yes/no. What that means for your data: Calibration.
- Games: 0.35 on the autoresearch games benchmark (0 = random play, 1 = expert), against −0.04. Near expert on ViZDoom basic, Atari Freeway, Acrobot and MountainCar; LunarLander is worse than random.
- Full scorecard: Scorecard: thaitea/laya-vision (201M).
The previous checkpoint¶
- Breadth. It averages 75% over 26 validation sets, with 93.7% on IconQA, 91.8% on DVQA and 89.9% on Hateful Memes. The full scorecard covers 34 validation sets, human-vote calibration, the games suite and latency: Scorecard.
scoreanswers. On held-out rubric data it scores 54% over 5 levels on VLFeedback response grading (prior-only baseline 27.5%). It is 0.8 levels off on average, against 1.4 for the baseline, and 0.38 levels off on 3-level damage severity: Score head results.
Evidence¶
Both rows' accuracies are backed by committed per-row predictions in
results/raw/. python benchmarks/verify_published.py recomputes them (and
calibrated ECE) and checks them against the same table in the README and against each checkpoint's metrics
(recommended,
previous). The rules for adding or changing a
published number are in AGENTS.md.
Other checkpoints¶
- thaitea/laya-vision-modernvbert-250m: ModernVBERT on 19 Cauldron subsets: results.
- thaitea/laya-vision-smolvlm-256m: the original, on A-OKVQA, ScienceQA and VQAv2 yes/no only.
Pin a checkpoint's revision when you load it (load_vlm(..., revision=...)): the id on the Hub can move, as
thaitea/laya-vision did on 2026-09-24.