Overview

FIRST (Faster Improved Listwise Reranking with Single Token Decoding) is a 7 billion parameter language model, fine-tuned from the HuggingFaceH4/zephyr-7b-beta model. It is specifically designed for listwise reranking, utilizing the output logits of the first generated identifier to directly produce a ranked ordering of candidates. The model undergoes single-stage fine-tuning on a converted alphabetic version of the RankZephyr dataset, which includes RankGPT-4 reorderings of OpenAI's Ada2 orderings for 5,000 queries.

Key Capabilities

• Efficient Listwise Reranking: Optimizes the ordering of candidate lists by directly leveraging output logits.
• Superior Performance: Demonstrates leading performance on various reranking datasets within the BEIR benchmark, surpassing models like Rank Vicuna and Rank Zephyr.
• English Language Support: Primarily trained and optimized for English reranking tasks.

Good for

• Search Engine Optimization: Enhancing the relevance and order of search results.
• Information Retrieval Systems: Improving the ranking of documents or passages based on query relevance.
• Recommendation Systems: Reordering recommended items to present the most relevant options first.

For more technical details and evaluation results, refer to the associated paper and the GitHub repository.