AdarshSingh7647/Eklav-Passage-8B
AdarshSingh7647/Eklav-Passage-8B is an 8 billion parameter Qwen3-based passage reranker model, developed by AdarshSingh7647, with a context length of 32768 tokens. It is specifically trained using the Eklav method, where the teacher's reasoning is provided as a hint and loss is applied only to the final answer. This model excels at determining the relevance of a passage to a given query, outperforming standard supervised fine-tuning (Std-SFT) methods on the BRIGHT nDCG@10 benchmark.
Loading preview...
Eklav-Passage-8B: An Advanced Passage Reranker
AdarshSingh7647/Eklav-Passage-8B is an 8 billion parameter passage reranker built upon the Qwen3 architecture. This model introduces a novel training methodology called Eklav, which leverages a teacher's reasoning as a hint within the prompt, with the loss function focused solely on the final answer. This approach significantly enhances its ability to accurately assess the relevance of passages to user queries.
Key Capabilities
- Superior Reranking Performance: Achieves a BRIGHT nDCG@10 score of 34.2, outperforming its Std-SFT counterpart (31.5) and other models like GLM-Z1-9B (33.2) and Qwen3-14B (34.7) in its Eklav variant.
- Pointwise Relevance Scoring: Functions as a pointwise reranker, similar to models like Rank1, by generating internal reasoning that concludes with a 'true' or 'false' relevance judgment.
- Efficient Inference: Designed for effective use with inference engines like vLLM, providing a structured method to score passage relevance based on the probability of the 'true' token.
- High Context Length: Supports a substantial context window of 32768 tokens, allowing for the evaluation of longer queries and passages.
Good For
- Improving Retrieval-Augmented Generation (RAG) Systems: Ideal for enhancing the precision of retrieved documents by accurately filtering and ranking passages based on their relevance to a query.
- Semantic Search Applications: Can be integrated into search engines to provide more relevant results by re-ranking initial search outputs.
- Question Answering Systems: Useful for selecting the most pertinent passages from a knowledge base to answer complex questions.
- Developers Seeking Advanced Reranking: Offers a robust solution for applications requiring high-fidelity passage relevance judgments, especially those looking to leverage innovative training techniques like Eklav.