AdarshSingh7647/TabRank

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

TabRank by AdarshSingh7647 is an 8 billion parameter Qwen3-based model fine-tuned for single-call generative listwise table reranking, capable of processing a question and multiple candidate tables in one prompt. It returns a full ranking without requiring pairwise scoring or cross-encoder passes. This model utilizes a reasoning-conditioned method, trained on distilled chain-of-thought traces, to achieve faster inference and superior generalization, particularly on out-of-distribution benchmarks. It excels at identifying the most relevant tables for a given query, making it ideal for information retrieval systems involving tabular data.

Loading preview...

TabRank: Single-Call Generative Table Reranking

TabRank is an 8 billion parameter model built on Qwen3, specifically fine-tuned for listwise table reranking in a single generative call. Unlike traditional methods that rely on pairwise comparisons or cross-encoders, TabRank processes a question and multiple candidate tables within one prompt, generating a complete ranked list directly. This approach significantly improves inference speed and efficiency.

Key Capabilities and Differentiators

  • Single-Call Reranking: Ranks multiple tables simultaneously, eliminating the need for iterative or pairwise scoring.
  • Reasoning-Conditioned Training: Trained on 6,728 chain-of-thought reasoning traces distilled from a teacher model. It learns to condition on this reasoning context to generate its own concise reasoning during inference, leading to better generalization.
  • Enhanced Generalization: Demonstrates strong performance on out-of-distribution benchmarks, outperforming its base model and standard SFT variants. It achieves a mean nDCG@10 of 0.720 across 12 benchmarks, winning 5 out of 7 out-of-distribution tests.
  • Faster Inference: Offers 4.3x faster inference compared to standard Chain-of-Thought (CoT) SFT methods.
  • Structured Output: Provides reasoning within a <think> block followed by a JSON object containing the ranked table indices.

Ideal Use Cases

  • Table-based Information Retrieval: Systems requiring efficient and accurate ranking of relevant tables for a given query.
  • Question Answering over Tables: Improving the precision of retrieved tables before extracting answers.
  • Data Discovery: Quickly identifying the most pertinent tables from a large corpus based on user queries.

Finetunes
6 models