AdarshSingh7647/Eklav-Passage-9B
AdarshSingh7647/Eklav-Passage-9B is a 9 billion parameter GLM-Z1-9B passage reranker model, fine-tuned using the Eklav method, which incorporates teacher's reasoning as a prompt hint. This model excels at determining the relevance of passages to a given query, achieving a BRIGHT nDCG@10 score of 33.2, outperforming its standard SFT counterpart. It functions as a pointwise reranker, generating internal reasoning before outputting a true/false relevance judgment, making it suitable for advanced information retrieval and search applications.
Loading preview...
Eklav-Passage-9B Overview
Eklav-Passage-9B is a 9 billion parameter GLM-Z1-9B model specifically designed for passage reranking. Developed by AdarshSingh7647, this model utilizes the innovative Eklav training method, where the teacher's reasoning is provided as a hint within the prompt, and the loss is calculated solely on the final answer. This approach enhances the model's ability to accurately assess passage relevance.
Key Capabilities
- Advanced Passage Reranking: Determines the relevance of a passage to a user's query with high accuracy.
- Eklav Training Method: Leverages teacher's reasoning during training to improve performance.
- Pointwise Reranker: Operates by generating internal reasoning (e.g.,
</think> trueor</think> false) to derive a relevance score. - Competitive Performance: Achieves a BRIGHT nDCG@10 score of 33.2, surpassing the standard SFT version (30.1) and outperforming Qwen3-4B and Qwen3-8B models in reranking tasks.
Good for
- Information Retrieval Systems: Enhancing the ranking of search results or retrieved documents.
- Question Answering Systems: Filtering and prioritizing relevant passages for answer generation.
- Semantic Search: Improving the precision of search queries by accurately identifying pertinent content.
- Research and Development: Exploring advanced fine-tuning techniques for specialized language models.