ali-vosoughi/oscar-llava-v1.5-13b-mixed-adapter

OSCaR LLaVA v1.5 13B Mixed Adapter

This repository contains the adapter artifact staged for the OSCaR public release.

Artifact Type

• Local staging directory: llava-v1.5-13b-lora-mixed-adapter
• Public repo id: ali-vosoughi/oscar-llava-v1.5-13b-mixed-adapter
• Training data condition: OSCaR plus the upstream LLaVA v1.5 mixed visual-instruction manifest (llava_final.json)

Files

• adapter_model.bin
• adapter_config.json
• config.json
• non_lora_trainables.bin

Loading

This is an adapter-only release. Pass --model-base with the matching Vicuna checkpoint when loading it.

Example:

python -m llava.serve.cli --model-path ali-vosoughi/oscar-llava-v1.5-13b-mixed-adapter --model-base lmsys/vicuna-13b-v1.5 --image-file /path/to/image.jpg

Training Configuration

• LLaVA v1.5 stack
• CLIP ViT-L/336 vision tower
• LoRA rank 128
• LoRA alpha 256
• learning rate 2e-4
• 1 epoch
• max sequence length 2048

Related Resources

• Code: https://github.com/nguyennm1024/OSCaR
• Dataset: https://huggingface.co/datasets/ali-vosoughi/oscar-dataset
• Paper: https://arxiv.org/abs/2402.17128