AdarshSingh7647/TabRankSingleTableNaive
AdarshSingh7647/TabRankSingleTableNaive is an 8 billion parameter generative listwise reranker built on Qwen3, designed for table retrieval tasks. This model efficiently orders candidate tables from most to least useful for answering a given question by processing all candidates in a single prompt. It is specifically trained on the NQ Tables dataset for single-table retrieval questions, outputting a direct ranking without explicit chain-of-thought reasoning for faster inference.
Loading preview...
TabRankSingleTableNaive Overview
AdarshSingh7647/TabRankSingleTableNaive is an 8 billion parameter model based on Qwen3, specifically engineered for generative listwise reranking in table retrieval. This variant, designated as "Single Table Answer Only," is optimized for speed by directly outputting the ranked list of tables without generating intermediate reasoning traces.
Key Capabilities & Features
- Single-Call Reranking: Processes a question and multiple candidate tables within a single prompt, returning a complete ranking.
- Direct Ranking Output: Generates a JSON object containing the ranked list of 1-based candidate positions, enabling quick integration into retrieval pipelines.
- Optimized for Inference Speed: Trained to output rankings directly, making it the fastest variant in the TabRank family for inference.
- Specialized Training: Fine-tuned using LoRA on the NQ Tables training split, focusing exclusively on single-table retrieval questions.
- Performance: Achieves strong results across various table question answering benchmarks, including SQA, TAT QA, HybridQA, and TabFact, with a mean recall@10 of 0.764 and MRR of 0.708.
Use Cases
- Table Question Answering (TQA): Ideal for systems that need to efficiently identify the most relevant tables from a set of candidates to answer a user's question.
- Information Retrieval: Enhances the precision of table retrieval systems by providing a refined ranking of search results.
- Single-Table Contexts: Best suited for scenarios where the answer is expected to be found within a single table, as it was trained specifically on single-table retrieval data.