AdarshSingh7647/TabRankMultiTableCoTGen
AdarshSingh7647/TabRankMultiTableCoTGen is an 8 billion parameter generative listwise reranker built on Qwen3, designed for table retrieval tasks. This model processes a question and multiple candidate tables in a single prompt, then outputs a ranked list of tables from most to least useful. It is specifically trained to generate a chain of thought alongside the ranking, making it suitable for use cases requiring explainability in table retrieval. The model is fine-tuned on a mix of NQ Tables and MultiTabQA datasets, excelling at multi-table reasoning.
Loading preview...
Overview
AdarshSingh7647/TabRankMultiTableCoTGen is an 8 billion parameter generative listwise reranker based on Qwen3, specialized in table retrieval. It takes a question and a set of candidate tables, then outputs a single, ordered list of tables ranked by their usefulness for answering the question. A key feature of this variant is its ability to generate a chain of thought (CoT) within a <think>...</think> block before producing the final ranking, providing explainability for its decisions.
Key Capabilities
- Generative Listwise Reranking: Reranks a list of candidate tables in a single generation. It reads all candidates in one prompt and emits the full ranking.
- Chain of Thought (CoT) Generation: Produces its own reasoning alongside the table ranking, useful for understanding the model's decision-making process.
- Multi-Table Reasoning: Trained on a combination of NQ Tables and MultiTabQA datasets, enhancing its performance on questions requiring information from multiple tables.
- Structured Output: Returns a JSON object containing the ranked list of one-based candidate positions.
Performance
Evaluated as a listwise reranker on a first-stage top 25 candidate list across four table question answering benchmarks (SQA, TAT QA, HybridQA, TabFact). It achieves a mean recall@10 of 0.782, ndcg@10 of 0.684, acc@10 of 0.624, and MRR of 0.745.
When to Use This Model
This model is ideal for applications requiring efficient and explainable table retrieval, particularly when dealing with questions that might span multiple tables. Its CoT generation makes it suitable for scenarios where transparency in ranking decisions is important.