AdarshSingh7647/TabRankStandardSFT
TabRankStandardSFT is an 8-billion parameter Qwen3-based model developed by Adarsh Singh, fine-tuned for single-call generative listwise table reranking with a 32768-token context length. It processes a question and multiple candidate tables in one prompt, generating a full ranking in a single output. This model utilizes chain-of-thought distillation, producing explicit reasoning traces to determine table relevance. It is designed for efficient table retrieval tasks where a ranked list of tables is required.
Loading preview...
TabRankStandardSFT: Generative Listwise Table Reranking
TabRankStandardSFT, developed by Adarsh Singh, is an 8-billion parameter model built on Qwen3, specifically fine-tuned for single-call generative listwise table reranking. Unlike traditional pairwise scoring methods, this model takes a question and a list of candidate tables within a single prompt and directly outputs a ranked list of tables.
Key Capabilities and Features
- Single-Call Reranking: Processes all candidate tables simultaneously to generate a complete ranking in one inference pass.
- Chain-of-Thought Distillation: Trained using chain-of-thought distillation from 6,728 teacher reasoning traces, enabling it to generate explicit
<think>...</think>reasoning blocks before outputting the final JSON ranking. - Output Format: Produces a JSON object containing one-indexed ranked table positions, allowing for direct mapping back to original table IDs.
- Context Length: Supports a substantial context length of 32768 tokens, accommodating multiple tables and detailed questions.
- Performance: Achieves a mean nDCG@10 of 0.700 across a mix of in-distribution and out-of-distribution benchmarks, positioning it between the base Qwen3-8B and the more advanced TabRank method.
Use Cases and Considerations
This model is ideal for applications requiring efficient and direct table relevance ranking, such as question answering over structured data or information retrieval systems. While it offers modest gains over the base model, its generalization to out-of-distribution data is less robust compared to the reasoning-conditioned TabRank method, and it has slower inference times. Developers should consider its balance of accuracy and inference speed for their specific table reranking needs.