AdarshSingh7647/Eklav-Passage-0.6B
AdarshSingh7647/Eklav-Passage-0.6B is a 0.8 billion parameter Qwen3-based passage reranker developed by AdarshSingh7647. It is 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 pointwise reranking, determining passage relevance to a query by generating reasoning and a true/false verdict, and achieves superior nDCG@10 scores compared to standard SFT models across various Qwen3 and GLM-Z1 base models.
Loading preview...
Eklav-Passage-0.6B Overview
Eklav-Passage-0.6B is a 0.8 billion parameter passage reranker built upon the Qwen3 architecture. Developed by AdarshSingh7647, this model utilizes a novel training approach called Eklav, which incorporates the teacher's reasoning as a prompt hint and applies loss solely on the final answer. This method enhances the model's ability to accurately assess the relevance of passages to a given query.
Key Capabilities
- Passage Reranking: Specializes in pointwise reranking, evaluating individual passages for relevance to a query.
- Reasoning Generation: The model generates its internal reasoning process before outputting a final true/false relevance judgment.
- Improved Performance: Demonstrates significantly higher nDCG@10 scores on the BRIGHT benchmark compared to standard Supervised Fine-Tuning (SFT) methods, outperforming Std-SFT versions across Qwen3-4B, Qwen3-8B, GLM-Z1-9B, and Qwen3-14B base models.
Good for
- Information Retrieval Systems: Ideal for improving the ranking of retrieved documents or passages in search engines and question-answering systems.
- Relevance Scoring: Provides a robust mechanism for scoring the relevance of text snippets to user queries.
- Applications requiring explainable relevance judgments: The model's ability to generate reasoning before its final verdict can be beneficial for understanding its decision-making process.