vismin-idefics2-8b

Model Details

The VisMin-Idefics2 model was developed as a fine-tuned version of the Idefics2 model, leveraging the VisMin dataset for enhanced performance in multimodal tasks. This model excels in visual-text alignment and is designed to handle tasks where models must differentiate between similar images based on textual descriptions. By employing the QLoRa technique and focusing on a rule-based selection of image-text pairs, the VisMin-Idefics2 model is optimized for fine-grained understanding and improved generalization across various multimodal benchmarks.

Model Summary

• Model Date: July 2024
• Model type: Multi-modal model (image+text)
• Parent Models: google/siglip-so400m-patch14-384 and mistralai/Mistral-7B-v0.1

Usage

This section shows snippets of code for generation for fine-tuned idefics2-8b. The codes only differ by the input formatting. Let's first define some common imports and inputs.

from transformers import AutoProcessor, AutoModelForVision2Seq

model_name_or_path = "path/to/fine-tuned-model"
if "A100" in gpu_name or "H100" in gpu_name:
     attn_implementation = "flash_attention_2"
else:
     attn_implementation = None

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)
processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b", do_image_splitting=False)
model = AutoModelForVision2Seq.from_pretrained(
    model_name_or_path,
    low_cpu_mem_usage=True,
    device_map="auto",
    torch_dtype=torch.float16,
    _attn_implementation=attn_implementation,  # only A100, H100 GPUs
    quantization_config=quantization_config
    if model_name_or_path in ["HuggingFaceM4/idefics2-8b", "HuggingFaceM4/idefics2-8b-base"]
    else None,
)

Bibtex

 @article{vismin2024,
    title={VisMin: Visual Minimal-Change Understanding},
    author={Awal, Rabiul and Ahmadi, Saba and Zhang, Le and Agrawal, Aishwarya},
    year={2024}
}