AdarshSingh7647/Eklav-0.6B-Reranker
AdarshSingh7647/Eklav-0.6B-Reranker is a 0.8 billion parameter reranker model based on the Qwen3-0.6B architecture. It utilizes the Eklav training method, which involves hint-conditioned supervised fine-tuning (SFT) where the model learns to continue a teacher's partial reasoning trace. This approach results in a reranker optimized for passage relevance tasks, demonstrating improved efficiency with 29% fewer training FLOPs compared to standard full trace CoT SFT, while maintaining competitive performance on benchmarks like BRIGHT.
Loading preview...
Eklav-0.6B-Reranker Overview
AdarshSingh7647/Eklav-0.6B-Reranker is a 0.8 billion parameter model built upon the Qwen3-0.6B base. Its core innovation lies in the Eklav training method, a form of hint-conditioned supervised fine-tuning (SFT). Instead of imitating a teacher's full reasoning trace, the model learns to complete a partial trace, conditioning its own reasoning on the provided hint.
Key Differentiators & Performance
- Efficient Training: Achieves a 29% reduction in training FLOPs compared to standard full trace CoT SFT, using the same base model and training data.
- Reranking Task: Specifically designed for passage reranking tasks, evaluated on benchmarks like BRIGHT and NevIR.
- Competitive Performance: While more efficient, it shows a modest performance difference of -11% on BRIGHT (nDCG@10, 12 domain average) compared to standard full trace CoT SFT.
- Pointwise Reranker: Functions as a pointwise reranker, similar to models like jhu-clsp/rank1-7b. It generates its own reasoning trace (e.g.,
</think> trueor</think> false) and scores relevance from the logits of the final token.
Use Cases
This model is ideal for applications requiring efficient and effective document or passage reranking, where the goal is to determine the relevance of a passage to a given query. Its training methodology allows for a more focused learning of reasoning continuation, making it suitable for integrating into retrieval-augmented generation (RAG) systems or search engines to improve result quality.