AdarshSingh7647/Eklav-4B-Reranker-CotGen

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

AdarshSingh7647/Eklav-4B-Reranker-CotGen is a 4 billion parameter language model based on Qwen3-4B, fine-tuned for passage reranking using a CotGen (Chain-of-Thought Generation) baseline. This model learns to continue reasoning from partial teacher traces to produce answers, rather than imitating full reasoning end-to-end. It is specifically designed as a pointwise reranker, excelling in tasks like BRIGHT and NevIR by generating reasoning traces that conclude with relevance scores.

Loading preview...

Overview

Eklav-4B-Reranker-CotGen is a 4 billion parameter model built upon the Qwen/Qwen3-4B base model. It is trained using a CotGen (Chain-of-Thought Generation) method, specifically a standard full trace CoT SFT baseline. The core idea behind Eklav is to train a student model to complete a teacher's reasoning process from a partial trace, rather than simply reproducing the entire reasoning sequence.

Key Capabilities

  • Passage Reranking: Optimized for tasks such as BRIGHT and NevIR, functioning as a pointwise reranker.
  • Reasoning Completion: Learns to continue and complete reasoning from incomplete teacher traces.
  • Relevance Scoring: Generates a reasoning trace ending in </think> true or </think> false, with relevance scored from the logits of the final token.

Performance

  • Achieves an average nDCG@10 of 29.3 on the BRIGHT dataset.

Usage Notes

  • This model is designed to be used with vLLM and a stop string (e.g., </think> true, </think> false) for optimal performance, as it can generate long reasoning traces before concluding. A fixed max_new_tokens cutoff with transformers.generate is not recommended due to potential for repetition without reaching a conclusion.