Reference¶

Factual reference material: the API, its inputs and outputs, the checkpoints, file formats and the published results. For learning and tasks, see the Tutorials and How-to guides.

API and checkpoints¶

• predict() and the answer schema: loading a checkpoint, the state, the three question types, every predict argument, and every field of the result.
• Checkpoints: the recommended checkpoint, its accuracy and calibration, and the evidence behind them.

Files and formats¶

• Atari training data format: the layout every Atari training source is written in.

Results¶

Generated from result files by scripts/eval_report.py, and not edited by hand:

• Scorecard: thaitea/laya-vision (201M): the recommended checkpoint on 34 validation sets, human-vote calibration, the games suite and latency.
• Scorecard: the previous checkpoint: the same suite on thaitea/laya-vision before 2026-09-24 (now laya-vision-smolvlm-256m-score).

Experiment reports:

• Score head results: the first runs with rubric-scored data, both backbones.
• ModernVBERT on The Cauldron: the bidirectional backbone post-trained on 19 Cauldron subsets.
• Image splitting on SmolVLM2: accuracy against latency at three split settings.
• Robustness: answers under meaning-preserving perturbations, and whether the image is used.
• Game training: ViZDoom and Atari policies, DAgger, and the cheaper preprocessing.
• Typed answers vs generated JSON: predict against a generative baseline on the same image.
• What didn't work: approaches that were tried and dropped.