AdarshSingh7647/Eklav-8B-Reranker
AdarshSingh7647/Eklav-8B-Reranker is an 8 billion parameter reranker model based on Qwen/Qwen3-8B, specifically trained using the Eklav method for passage reranking tasks. This model learns to continue reasoning from partial traces, achieving a +9% improvement on BRIGHT (nDCG@10) compared to standard full trace CoT SFT while reducing training FLOPs by 32%. It functions as a pointwise reranker, generating its own reasoning to determine relevance for a given query and passage.
Loading preview...
Eklav-8B-Reranker Overview
Eklav-8B-Reranker is an 8 billion parameter model developed by AdarshSingh7647, built upon the Qwen/Qwen3-8B base model. It is specifically designed for passage reranking tasks, utilizing a novel training method called Eklav. Unlike traditional methods that imitate a teacher's full reasoning trace, Eklav trains the model to continue reasoning from a partial trace, making it more efficient and effective.
Key Capabilities & Performance
- Enhanced Reranking Performance: Achieves a +9% improvement in nDCG@10 on the BRIGHT benchmark (12-domain average) compared to standard full trace CoT SFT, using the same base model and training data.
- Training Efficiency: Demonstrates a significant 32% reduction in training FLOPs compared to standard full trace CoT SFT.
- Pointwise Reranking: Operates as a pointwise reranker, similar to models like Rank1, by generating its own reasoning trace (e.g.,
</think> trueor</think> false) to score relevance based on final token logits. - Autonomous Reasoning: At inference time, the model reasons independently from a bare prompt without requiring teacher hints, making it practical for real-world applications.
Use Cases
- Information Retrieval Systems: Ideal for improving the relevance ranking of search results or retrieved documents.
- Question Answering Systems: Can be integrated to rerank passages before answer extraction, enhancing accuracy.
- Content Recommendation: Useful for ranking content based on user queries or preferences.