AdarshSingh7647/Eklav-14B-Reranker-CotGen
AdarshSingh7647/Eklav-14B-Reranker-CotGen is a 14 billion parameter language model based on Qwen/Qwen3-14B, specifically fine-tuned for passage reranking tasks. It utilizes a CotGen (Chain-of-Thought Generation) baseline, a standard full trace CoT SFT method, to measure improvements in reasoning. This model excels at determining passage relevance by generating a reasoning trace and scoring based on final token logits, achieving an average nDCG@10 of 32.5 on the BRIGHT dataset.
Loading preview...
Overview
AdarshSingh7647/Eklav-14B-Reranker-CotGen is a 14 billion parameter model built upon the Qwen/Qwen3-14B architecture, designed for passage reranking. It employs a CotGen (Chain-of-Thought Generation) baseline, which is a standard full trace CoT Supervised Fine-Tuning (SFT) method. This model's training objective focuses on learning to continue reasoning from a partial teacher's trace rather than imitating an end-to-end reasoning process.
Key Capabilities
- Passage Reranking: Specialized in determining the relevance of passages to a given query.
- Reasoning Trace Generation: Generates a detailed reasoning trace before outputting a relevance decision.
- Logit-based Scoring: Relevance is scored from the logits of the final
trueorfalsetoken, avoiding reliance on parsing generated text. - Performance: Achieves an average nDCG@10 of 32.5 on the BRIGHT dataset for reranking tasks.
When to Use This Model
- Pointwise Reranking: Ideal for applications requiring pointwise reranking, similar to models like jhu-clsp/rank1-7b.
- Complex Relevance Judgments: Suitable for scenarios where a reasoning trace can aid in more accurate relevance determination.
- vLLM Integration: Recommended for use with vLLM and stop-string based sampling to handle potentially long reasoning traces efficiently and prevent degeneration.