| { | |
| "model_id": "openvla-micro", | |
| "description": "Small-vision VLA trained on LIBERO-90 for CPU robot deployment", | |
| "vision": { | |
| "backbone": "dinosiglip-vit-s-b-224px", | |
| "dino_model": "vit_small_patch14_reg4_dinov2.lvd142m", | |
| "dino_dim": 384, | |
| "dino_patches": 256, | |
| "siglip_model": "vit_base_patch16_siglip_224", | |
| "siglip_dim": 768, | |
| "siglip_patches": 196, | |
| "total_patches": 452, | |
| "total_embed_dim": 1152, | |
| "image_size": 224 | |
| }, | |
| "projector": { | |
| "shim_hidden": 2048, | |
| "shim_out_dim": 8704, | |
| "proj2_in": 8704, | |
| "proj2_out": 896, | |
| "proj4_in": 896, | |
| "proj4_out": 896, | |
| "trainable_params": 38107938 | |
| }, | |
| "llm": { | |
| "model_id": "qwen25-0_5b-extra", | |
| "hf_id": "Qwen/Qwen2.5-0.5B", | |
| "vocab_size": 151936, | |
| "hidden_dim": 896, | |
| "extra_tokens": 256 | |
| }, | |
| "action": { | |
| "dof": 7, | |
| "normalization": "minmax_q99", | |
| "tokenizer_bins": 256, | |
| "dataset": "libero_90" | |
| }, | |
| "training": { | |
| "optimizer_steps": 5000, | |
| "effective_batch_size": 64, | |
| "micro_batch": 8, | |
| "gradient_accumulation": 8, | |
| "learning_rate": 0.0002, | |
| "lr_schedule": "200_warmup_cosine_to_1e-5", | |
| "lora_rank": 8, | |
| "checkpoint_source": "MiniVLA" | |
| } | |
| } | |