laya-vision

Non-autoregressive System 1 decision engine with vision: typed decisions (choice / score / noul) over text + up to 4 images in one forward pass. Fine-tuned from convaiinnovations/laya with a google/siglip2-base-patch16-224 tower (fp32, overflow-safe) fused as 32 prefix tokens per image, trained with RLCD (proper-scoring-rule policy gradient + soft cross-entropy).

Training

• Data: ScienceQA + Flickr30k verification, text-dropout fallback
• Epochs: 3 | micro-batch 4 x2 GPUs x2 accum (effective 16)
• LR: encoder 2.5e-05, vision 3e-05, head 0.0001 (AdamW, cosine)
• Final train loss: 0.1842 | held-out acc: 0.198 | ECE: 0.000
• Calibration temperatures (choice, score, noul): [nan, 1.2, nan]

Use

# inference support lands in laya/ after this checkpoint publishes;
# checkpoint layout: model.safetensors, rl_agent_config.json,
# encoder/, tokenizer/, image_processor/