YuchengShi/LLaVA-v1.5-7B-CUB-200

Fine-Grained Visual Classification on CUB-200

Project Page: SelfSynthX.

Paper on arXiv: Enhancing Cognition and Explainability of Multimodal Foundation Models with Self-Synthesized Data

This model is a fine-tuned multimodal foundation model developed on the LLaVA-1.5-7B-hf base, optimized for fine-grained visual classification and explainability using the CUB-200 dataset.

Key Details

• Base Model: LLaVA-1.5-7B
• Dataset: CUB-200 (Caltech-UCSD Birds-200-2011)
• Innovation:
  • Self-Synthesized Data: Generates interpretable explanations by extracting image-specific visual concepts using the Information Bottleneck principle.
  • Iterative Fine-Tuning: Uses reward model-free rejection sampling to progressively improve classification accuracy and explanation quality.
• Intended Use: Fine-grained bird species identification with human-verifiable explanations.

How to Use

import requests
from PIL import Image
import torch
from transformers import AutoProcessor, LlavaForConditionalGeneration

model_id = "YuchengShi/LLaVA-v1.5-7B-CUB-200"
model = LlavaForConditionalGeneration.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    low_cpu_mem_usage=True,
).to("cuda")
processor = AutoProcessor.from_pretrained(model_id)

conversation = [
    {
      "role": "user",
      "content": [
          {"type": "text", "text": "What are these?"},
          {"type": "image"},
        ],
    },
]
prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
image_file = "cub-200/test1.png"
raw_image = Image.open(requests.get(image_file, stream=True).raw)
inputs = processor(images=raw_image, text=prompt, return_tensors='pt').to("cuda", torch.float16)

output = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(processor.decode(output[0][2:], skip_special_tokens=True))

Training & Evaluation

• Training: Fine-tuned using LoRA on CUB-200 with iterative rejection sampling.
• Evaluation: Demonstrates higher accuracy and robust, interpretable explanations compared to baseline models.

Citation

If you use this model, please cite:

@inproceedings{
  shi2025enhancing,
  title={Enhancing Cognition and Explainability of Multimodal Foundation Models with Self-Synthesized Data},
  author={Yucheng Shi and Quanzheng Li and Jin Sun and Xiang Li and Ninghao Liu},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=lHbLpwbEyt}
}