AdarshSingh7647/Eklav-14B-Reranker
AdarshSingh7647/Eklav-14B-Reranker is a 14 billion parameter reranker model based on Qwen3-14B, trained using the Eklav method for hint-conditioned supervised fine-tuning. This model excels at passage reranking, demonstrating a +7% improvement on the BRIGHT benchmark (nDCG@10) compared to standard full trace CoT SFT. It is optimized for efficient inference as a pointwise reranker, generating its own reasoning trace to determine relevance.
Loading preview...
Eklav-14B-Reranker Overview
AdarshSingh7647/Eklav-14B-Reranker is a 14 billion parameter model built upon the Qwen3-14B base architecture. It is specifically designed for passage reranking tasks, utilizing a novel training approach called Eklav (hint-conditioned SFT).
Key Differentiators & Performance
Unlike traditional Chain-of-Thought (CoT) distillation, the Eklav method trains the model to continue a teacher's partial reasoning trace rather than imitating it end-to-end. This approach yields significant benefits:
- Improved Reranking Performance: Achieves a +7% increase 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.
- Reduced Training Cost: Demonstrates a 24% reduction in training FLOPs compared to standard full trace CoT SFT.
Use as a Reranker
This model functions as a pointwise reranker, similar to jhu-clsp/rank1-7b. It generates an internal reasoning trace, concluding with either true or false to indicate relevance. Relevance scores are derived from the logits of these final tokens. During inference, the model operates without a teacher hint, reasoning independently from the provided query and passage. The model achieves an average nDCG@10 of 34.7 on the BRIGHT benchmark.