metadata

datasets:
  - liuhaotian/LLaVA-Instruct-150K
  - liuhaotian/LLaVA-CC3M-Pretrain-595K
language:
  - en
metrics:
  - accuracy
pipeline_tag: visual-question-answering

DinoV2-SigLIP-Phi3(LoRA) VLM

Vision Encoder - DinoV2 + SigLIP @384px resolution. Why 2 vision encoders?
Connector - MLP (Dino and SigLIP features are concatenated and then projected to Phi3 representation space)
Language Model - Phi3 + LoRA
Pre-train (Align) Dataset - LLaVA-CC3M-Pretrain-595K
Fine-tune (Instruction) Dataset - LLAVA-v1.5-Instruct + LRV-Instruct

Scripts to build and train the models are available at NMS05/DinoV2-SigLIP-Phi3-LoRA-VLM.