AdarshSingh7647/Eklav-4B-Reranker
Eklav-4B-Reranker by AdarshSingh7647 is a 4 billion parameter passage reranking model based on Qwen/Qwen3-4B, specifically trained using the Eklav method (hint-conditioned SFT). This approach allows the model to learn to continue reasoning from partial teacher traces, achieving an 8% improvement on BRIGHT (nDCG@10) compared to standard CoT SFT while reducing training FLOPs by 31%. It functions as a pointwise reranker, generating its own reasoning to determine passage relevance for a given query.
Loading preview...
Eklav-4B-Reranker: Efficient Passage Reranking
Eklav-4B-Reranker is a 4 billion parameter model developed by AdarshSingh7647, built upon the Qwen/Qwen3-4B base. Its core innovation lies in its training methodology, known as Eklav (hint-conditioned SFT). Instead of fully imitating a teacher's reasoning, the model learns to complete reasoning from partial traces, fostering independent reasoning capabilities.
Key Capabilities and Performance
- Enhanced Reranking Performance: Achieves an 8% improvement on the BRIGHT benchmark (nDCG@10, 12-domain average) compared to traditional full-trace Chain-of-Thought (CoT) Supervised Fine-Tuning (SFT) using the same base model and training data.
- Training Efficiency: Demonstrates a 31% reduction in training FLOPs compared to standard full-trace CoT SFT, making it more resource-efficient.
- Pointwise Reranking: Operates as a pointwise reranker, similar to models like Rank1. It generates an internal reasoning trace and scores relevance based on the logits of a final 'true' or 'false' token, rather than parsing generated text.
- Autonomous Reasoning: During inference, the model reasons independently from a bare prompt without any teacher hints, reflecting its ability to generate relevant judgments on its own.
Use Cases
This model is ideal for applications requiring efficient and accurate passage reranking, such as:
- Improving the relevance of search results in information retrieval systems.
- Filtering and ranking documents based on query relevance.
- Enhancing retrieval-augmented generation (RAG) pipelines by providing better context to LLMs.