AdarshSingh7647/Eklav-9B-Reranker
AdarshSingh7647/Eklav-9B-Reranker is a 9 billion parameter language model based on the zai-org/GLM-Z1-9B-0414 architecture, specifically trained as a pointwise passage reranker. It utilizes the Eklav training method, which involves hint-conditioned supervised fine-tuning (SFT) where the model learns to continue partial reasoning traces. This approach allows it to achieve a +10% improvement on the BRIGHT benchmark (nDCG@10) compared to standard full trace CoT SFT, while also reducing training FLOPs by 30%. It is designed for tasks requiring relevance scoring, such as information retrieval and search result ranking.
Loading preview...
Eklav-9B-Reranker: An Efficient Passage Reranker
AdarshSingh7647/Eklav-9B-Reranker is a 9 billion parameter model built upon the zai-org/GLM-Z1-9B-0414 base, specialized for passage reranking. Its core innovation lies in the Eklav training method, which teaches the model to complete reasoning traces from partial hints, rather than simply imitating full teacher reasoning end-to-end. This method significantly enhances its performance and efficiency.
Key Capabilities & Differentiators
- Superior Reranking Performance: Achieves a +10% improvement on the BRIGHT benchmark (nDCG@10, 12 domain average) compared to traditional full trace CoT SFT, using the same base model and training data.
- Reduced Training Cost: Demonstrates a 30% reduction in training FLOPs compared to standard full trace CoT SFT, making it more resource-efficient.
- Pointwise Reranking: Functions as a pointwise reranker, similar to
Rank1, where relevance is determined by the model generating a reasoning trace ending intrueorfalse, and scoring based on the logits of these final tokens. - Hint-Conditioned SFT: Learns to reason and produce answers by conditioning its own reasoning on partial teacher traces during training, leading to more robust inference without hints.
Ideal Use Cases
- Information Retrieval Systems: Enhancing the relevance of search results by accurately reranking passages based on a query.
- Question Answering Systems: Improving the selection of relevant documents or snippets to answer user queries.
- Any application requiring precise passage relevance scoring.
This model is particularly suited for developers looking for an efficient and high-performing reranker that leverages a novel training paradigm to surpass conventional distillation methods.