AdarshSingh7647/TabRankNaive
AdarshSingh7647/TabRankNaive is an 8 billion parameter Qwen3-based model fine-tuned for single-call generative listwise table reranking. It processes a question and multiple candidate tables in one prompt, returning a full ranking as a JSON object. This 'Answer-Only' variant is optimized for inference speed by directly outputting the ranking without a reasoning trace, making it suitable for efficient table relevance tasks.
Loading preview...
TabRankNaive: Efficient Listwise Table Reranking
AdarshSingh7647/TabRankNaive is an 8 billion parameter model built on Qwen3-8B, specifically fine-tuned for single-call generative listwise table reranking. Unlike traditional methods that might involve pairwise scoring or cross-encoder passes, this model takes a question and a list of candidate tables within a single prompt and directly generates a ranked list of tables.
Key Capabilities
- Generative Listwise Reranking: Ranks multiple candidate tables in one go, providing a complete ordered list.
- Answer-Only Variant: This specific checkpoint is fine-tuned to output the ranking directly as a JSON object, without including a reasoning trace, which optimizes it for faster inference.
- Input/Output Format: Accepts a chat message with a question followed by labeled candidate tables and outputs a JSON object containing one-indexed ranked table positions.
- Training Data: Trained on a mix of NQ Tables and MultiTabQA datasets.
Good for
- Fast Table Relevance: Ideal for applications requiring quick and efficient ranking of tables based on a given query.
- Direct Ranking Output: Suitable for systems that consume a structured JSON output for table ordering.
- Integration with LLM Workflows: Designed to be used with frameworks like vLLM or Hugging Face Transformers for seamless integration into existing LLM pipelines.
For more detailed comparisons and verified evaluation results, users are directed to related models like TabRank (reasoning-conditioned) and TabRankStandardSFT.