AdarshSingh7647/Eklav-Passage-14B

Eklav-Passage-14B Overview

Eklav-Passage-14B is a 14 billion parameter passage reranker built upon the Qwen3 architecture, developed by AdarshSingh7647. This model is specifically designed to enhance the relevance of retrieved passages for a given query, utilizing a unique training methodology called Eklav. The Eklav method incorporates the teacher's reasoning as a hint during training, with the loss function focused solely on the final answer, leading to improved performance in reranking tasks.

Key Capabilities

• Advanced Passage Reranking: Specializes in pointwise reranking, similar to the Rank1 approach, by generating internal reasoning before outputting a 'true' or 'false' relevance judgment.
• Eklav Training Advantage: Leverages a novel training paradigm that provides reasoning hints, resulting in higher accuracy in determining passage relevance.
• Superior Performance: Achieves notable improvements in nDCG@10 scores across various Qwen3 and GLM-Z1 models, with Eklav-Passage-14B (Qwen3-14B) reaching 34.7 nDCG@10, outperforming its Std-SFT counterpart (32.5).
• High Context Length: Supports a substantial context window of 32768 tokens, allowing for the processing of longer queries and passages.

Good For

• Improving Retrieval-Augmented Generation (RAG) Systems: Enhances the quality of retrieved documents, leading to more accurate and relevant responses from LLMs.
• Search Engine Optimization: Can be integrated into search pipelines to re-rank initial search results, presenting users with more pertinent information.
• Information Retrieval Tasks: Ideal for any application requiring precise identification of relevant passages from a larger corpus.
• Applications requiring explainable relevance judgments: The model's internal reasoning process (ending in </think> true or </think> false) can offer insights into its relevance decisions.