AdarshSingh7647/Eklav-Passage-4B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 26, 2026Architecture:Transformer Featherless Exclusive Cold

Eklav-Passage-4B is a 4 billion parameter passage reranker developed by AdarshSingh7647, built upon the Qwen3 architecture. It is specifically trained using the Eklav method, which incorporates teacher's reasoning as a prompt hint and optimizes loss solely on the final answer. This approach significantly enhances its reranking capabilities, achieving a BRIGHT nDCG@10 score of 31.5, outperforming standard SFT versions and larger models like Qwen3-8B and GLM-Z1-9B in its class. Its primary use case is as a pointwise reranker for information retrieval, determining passage relevance to a query by generating reasoning and a true/false conclusion.

Loading preview...

Eklav-Passage-4B Overview

Eklav-Passage-4B is a 4 billion parameter passage reranker based on the Qwen3 architecture, developed by AdarshSingh7647. This model distinguishes itself through its unique training methodology, named Eklav. The Eklav method integrates a teacher's reasoning directly into the prompt as a hint, with the loss function focused exclusively on the final answer. This technique has demonstrated superior performance in passage reranking tasks.

Key Capabilities

  • Enhanced Reranking Performance: Achieves a BRIGHT nDCG@10 score of 31.5, surpassing its Std-SFT counterpart (29.3) and even larger models like Qwen3-8B (31.5 for Std-SFT) and GLM-Z1-9B (30.1 for Std-SFT) in its class.
  • Reasoning-Based Relevance: Operates as a pointwise reranker, similar to Rank1, where it generates internal reasoning before concluding with a true or false relevance judgment. The probability of true at the final token indicates relevance.
  • Efficient for Retrieval: Designed to effectively filter and order passages based on their relevance to a given query, making it suitable for search and question-answering systems.

Good For

  • Information Retrieval Systems: Ideal for improving the quality of search results by re-ranking retrieved passages.
  • Question Answering: Can be used to select the most relevant passages from a corpus to answer a user's question.
  • Applications Requiring Contextual Relevance: Any system that benefits from precise identification of relevant textual content based on a query.

Finetunes
1 models