quicktensor/blockrank-msmarco-mistral-7b

BlockRank-Mistral-7B: Scalable In-context Ranking with Generative Models

BlockRank-Mistral-7B is a fine-tuned version of Mistral-7B-Instruct-v0.3 optimized for efficient in-context document ranking. It implements BlockRank, a method that makes LLMs efficient and scalable for ranking by aligning their internal attention mechanisms with the structure of the ranking task.

Key Features

• Linear Complexity Attention: Structured sparse attention reduces complexity from O(n²) to O(n)
• 2-4× Faster Inference: Attention-based scoring eliminates autoregressive decoding
• Auxiliary Contrastive Loss: Mid-layer contrastive objective improves relevance signals
• Strong Zero-shot Generalization: SOTA performance on BEIR benchmarks

Citation

If you use this model, please cite:

@article{gupta2025blockrank,
  title={Scalable In-context Ranking with Generative Models},
  author={Gupta, Nilesh and You, Chong and Bhojanapalli, Srinadh and Kumar, Sanjiv and Dhillon, Inderjit and Yu, Felix},
  journal={arXiv preprint arXiv:2510.05396},
  year={2025}
}

Model Card Contact

For questions or issues, please open an issue on GitHub.

Additional Resources

• Paper: arXiv:2510.05396
• Code: GitHub Repository
• Dataset: HuggingFace Dataset
• Demo: Colab Notebook

License

This model is released under the MIT License. See LICENSE for details.