AdarshSingh7647/Eklav-9B-Reranker-CotGen
AdarshSingh7647/Eklav-9B-Reranker-CotGen is a 9 billion parameter language model based on the GLM-Z1 architecture, specifically fine-tuned for passage reranking tasks. This model utilizes a standard full trace CoT (Chain-of-Thought) SFT baseline, learning to continue reasoning from partial teacher traces rather than direct imitation. It is optimized for pointwise reranking, demonstrated by its performance on benchmarks like BRIGHT (30.1 nDCG@10 average), making it suitable for improving search relevance and information retrieval systems.
Loading preview...
Eklav-9B-Reranker-CotGen Overview
This model, developed by AdarshSingh7647, is a 9 billion parameter reranker built upon the zai-org/GLM-Z1-9B-0414 base model. It employs a novel training methodology called CotGen, which focuses on teaching the model to extend a teacher's partial reasoning trace rather than merely replicating full reasoning paths. This approach aims to foster more independent and robust reasoning capabilities in the student model.
Key Capabilities
- Passage Reranking: Specifically designed and trained for passage reranking tasks, improving the relevance of retrieved documents.
- Chain-of-Thought (CoT) Distillation: Utilizes a standard full trace CoT SFT baseline, serving as a benchmark for the Eklav training method's improvements.
- Pointwise Reranking: Functions as a pointwise reranker, similar to models like Rank1, where relevance is determined by logits of
trueorfalsetokens after a reasoning trace. - Performance: Achieves an average nDCG@10 of 30.1 on the BRIGHT benchmark, indicating its effectiveness in ranking relevant passages.
Usage Recommendations
For optimal performance, especially given its potential for long reasoning traces, it is recommended to use this model with vLLM and a stop string (e.g., </think> true, </think> false) rather than a fixed max_new_tokens limit. This prevents the model from degenerating into repetition if it exceeds a token budget before completing its reasoning.