AdarshSingh7647/Eklav-Table-8B
AdarshSingh7647/Eklav-Table-8B is an 8 billion parameter Qwen3-based model with a 32768 token context length, specifically fine-tuned for single-call listwise table reranking. It utilizes the Eklav training method, where the model learns to generate a full ranking of candidate tables for a given question in one go, incorporating teacher's reasoning during training. This model excels at identifying the most relevant tables for a query, outperforming base Qwen3-8B and other SFT variants on various table-related benchmarks like SQA, TAT-QA, and TabFact.
Loading preview...
Overview
AdarshSingh7647/Eklav-Table-8B is an 8 billion parameter language model built upon Qwen3-8B, specifically designed for single-call listwise table reranking. This model's core innovation lies in its training methodology, named Eklav, which involves providing the teacher's reasoning as context during training while optimizing for the final ranking output. At inference, the model generates its own concise reasoning before producing the ranked list of tables.
Key Capabilities
- Single-call listwise table reranking: Given a question and multiple candidate tables, the model outputs a full ranking in a single generation step.
- Reasoning generation: It produces a
<think>block containing its reasoning process before outputting the final ranked list. - High performance on table-related tasks: Eklav-Table-8B demonstrates superior performance compared to the base Qwen3-8B and other SFT variants across various benchmarks, including SQA, TAT-QA, TabFact, NQ-Tables, AITQA, FeTaQA, OTT-QA, and OpenWikiTables.
Good for
- Applications requiring efficient and accurate ranking of tabular data in response to natural language queries.
- Information retrieval systems where identifying the most relevant tables from a large set is crucial.
- Building intelligent agents that can understand and process structured data effectively.