AdarshSingh7647/Std-SFT-Passage-4B
AdarshSingh7647/Std-SFT-Passage-4B is a 4 billion parameter Qwen3-based passage reranker model, developed by AdarshSingh7647, with a context length of 32768 tokens. It is trained using the Std-SFT method, which focuses the loss on the teacher's reasoning and final answer. This model functions as a pointwise reranker, determining passage relevance to a query by outputting 'true' or 'false' after its reasoning, making it suitable for information retrieval and search applications.
Loading preview...
Overview
AdarshSingh7647/Std-SFT-Passage-4B is a 4 billion parameter passage reranker built on the Qwen3 architecture. Developed by AdarshSingh7647, this model is specifically designed for information retrieval tasks, aiming to determine the relevance of a given passage to a query. It utilizes a unique training methodology called Std-SFT, where the loss function is applied to both the teacher's reasoning process and the final answer, enhancing its ability to make accurate relevance judgments.
Key Capabilities
- Passage Reranking: Efficiently reorders passages based on their relevance to a user query.
- Pointwise Relevance Scoring: Operates as a pointwise reranker, similar to models like Rank1, by generating internal reasoning before outputting a 'true' or 'false' relevance judgment.
- Reasoning-based Output: The model explicitly generates its reasoning process, ending with
</think> trueor</think> false, allowing for a probability-based relevance score derived from these tokens.
Good for
- Improving Search Accuracy: Enhancing the precision of search results by re-ranking retrieved documents.
- Information Retrieval Systems: Integrating into systems that require fine-grained relevance scoring for passages.
- Research and Development: Exploring the impact of Std-SFT training on reranking performance, especially when compared to other methods like Eklav.