Dauka-transformers/Compact_VLM_filter

Compact VLM Filter: Image-caption filtration-oriented Qwen2VL model

This model is a fine-tuned version of Qwen/Qwen2-VL-2B-Instruct trained to perform filtration-oriented image-text evaluation, based on our custom dataset.

🔍 Intended Use

The model is designed to:

• Evaluate alignment of image and caption
• Provide image/caption alignment scores and textual justification for noisy web-scale data
• Supports local deployment for cost-efficient training data filtration

🏋️ Training Details

• Base model: Qwen/Qwen2-VL-2B-Instruct
• Fine-tuning objective: in-context evaluation of aligment, quality and safety
• Dataset: ~4.8K samples with score, justification, caption, and image

🤝 Acknowledgements

Thanks to the Qwen team for open-sourcing their VLM models, which serve as the foundation for our filtration-oriented model.

📜 License

Licensed under the Apache License 2.0.