AdarshSingh7647/TabRankMultiTableCoTGen
AdarshSingh7647/TabRankStandardSFT is an 8-billion parameter Qwen3-based model fine-tuned for single-call generative listwise table reranking, capable of processing multiple candidate tables in one prompt. This variant uses standard chain-of-thought distillation, generating its own reasoning trace before outputting a ranked list of tables. It is trained on NQ Tables and MultiTabQA datasets, excelling at identifying relevant tables for a given question in a single pass.
Loading preview...
TabRankStandardSFT: Generative Table Reranking
AdarshSingh7647/TabRankStandardSFT is an 8-billion parameter model built on Qwen3-8B, designed for single-call generative listwise table reranking. Unlike traditional pairwise scoring methods, this model processes a question and multiple candidate tables within a single prompt, generating a full ranking in one output.
Key Capabilities and Features
- Single-Call Reranking: Ranks multiple tables simultaneously, improving efficiency by eliminating iterative pairwise comparisons.
- Chain-of-Thought (CoT) Distillation: Employs standard CoT distillation, where the model generates an explicit
<think>...</think>reasoning block before producing the final ranked list. This variant is trained with loss over the full reasoning trace. - Multi-Table Retrieval: Trained on a mix of NQ Tables and MultiTabQA datasets, enabling it to handle both single and multi-table retrieval questions effectively.
- Output Format: Provides a JSON object containing one-indexed ranked table positions, allowing for easy integration into retrieval pipelines.
Performance and Use Cases
This Standard SFT variant achieves a mean nDCG@10 of 0.700 across 12 benchmarks, including both in-distribution (SQA, TAT-QA, HybridQA, TabFact, NQ-Tables) and out-of-distribution datasets (OpenWikiTables, OTT-QA, MultiHiertt, AIT-QA, FeTaQA, StatCanDialogue, WatsonxDocsQA). While it shows modest gains over the base Qwen3-8B, its generalization is noted to be weaker than the reasoning-conditioned TabRank method (e.g., TabRankMultiTableCoTCond).
Use this model if you need:
- An efficient, single-pass solution for ranking candidate tables based on relevance to a query.
- A model that provides an interpretable reasoning trace alongside its ranking decision.
- To rerank tables in applications involving complex question answering or information retrieval from structured data.