MLM-Filter-llama-3.2-3b Model Card

Model details

Model type: MLM-Filter-llama-3.2-3b is an open-source MLLM trained to assess the data quality of image-text paired data. It can generate 4 quality metrics for image-text data: Image Text Matching, Object Detail Fulfillment, Caption Text Quality, and Semantic Understanding.

Model date: MLM-Filter-llama-3.2-3b was trained in Oct 2024.

Paper or resources for more information: https://mlm-filter.github.io/

@article{wang2024finetuned,
  title={Finetuned Multimodal Language Models Are High-Quality Image-Text Data Filters},
  author={Wang, Weizhi and Mrini, Khalil and Yang, Linjie and Kumar, Sateesh and Tian, Yu and Yan, Xifeng and Wang, Heng},
  journal={arXiv preprint arXiv:2403.02677},
  year={2024}
}

License

Llama 3 is licensed under the LLAMA 3 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved.

Where to send questions or comments about the model: https://github.com/Victorwz/MLM_Filter/issues

Intended use

Primary intended uses: MLM-Filter can be used as a drop-in replacement for CLIPScore in these tasks:

1. Score image-text data in large-scale pre-training dataset and then filter high-quality subsets based on the scores (For training MLLMs or VLMs, please consider to jointly use the Image-Text Matching score and the Object Detail Fulfillment score);

2. Evaluate the image-text alignment for image2text or text2image generation models;

3. Any potential applications with the need to calculate the image-text alignment.

Training dataset

• 665k instruction sampled from LLaVA-1.5 665k data.
• 4k instructions on image-text data quality assessment tasks ranging across 4 metrics.

Usage Sample

Please follow the instructions in https://github.com/Victorwz/MLM_Filter.