AdarshSingh7647/TabRankSingleTableCoTGen

TabRankSingleTableCoTGen Overview

This model, built on the Qwen3 8B architecture, functions as a generative listwise reranker for table retrieval. Its primary purpose is to take a natural language question and a set of candidate tables, then output a ranked list of these tables based on their relevance to the question. A key differentiator of this specific variant is its ability to generate a chain of thought (CoT) explanation within a <think>...</think> block before producing the final ranking, making it suitable for use cases requiring transparency in the ranking process.

Key Capabilities

• Single-Call Reranking: Processes all candidate tables in a single prompt and generates the full ranking in one generation.
• Chain of Thought Generation: Provides an explicit reasoning process alongside the ranked output, enhancing interpretability.
• Table Relevance Ranking: Orders candidate tables from most to least useful for answering a given question.
• JSON Output: Returns rankings in a structured JSON format, mapping candidate positions to a reranked list.

Training and Performance

The model was trained exclusively on the NQ Tables training split, focusing on single-table retrieval questions. It achieves strong performance across various table question answering benchmarks, including SQA, TAT QA, HybridQA, and TabFact. For instance, it shows a mean recall@10 of 0.764 and a mean MRR of 0.695 when reranking a first-stage top 25 candidate list.

Good for

• Applications requiring explainable table retrieval where a reasoning trace is beneficial.
• Scenarios involving single-table question answering where precise table ranking is crucial.
• Integrating into systems that can leverage a generative reranker for improved table relevance.