dnaihao/mistral-v0.3-tablebench
dnaihao/mistral-v0.3-tablebench is a 7 billion parameter instruction-tuned language model, a replication of TableBenchLLM, built upon Mistral-7B-Instruct-v0.3. Developed by dnaihao, this model is specifically fine-tuned on the TableBench instruction-tuning corpus to analyze the impact of base models versus training data in table-related tasks. It is primarily designed for research in understanding the factors influencing performance in Table LLMs, demonstrating that the base model choice is more dominant than the training data itself.
Loading preview...
Overview
dnaihao/mistral-v0.3-tablebench is a 7 billion parameter instruction-tuned model, replicating the TableBenchLLM, and is built on the Mistral-7B-Instruct-v0.3 base model. This model was developed by dnaihao as part of an EACL 2026 Findings paper titled "What Really Matters for Table LLMs? A Meta-Evaluation of Model and Data Effects." The research systematically investigates the contributions of the base model versus the training data by fine-tuning three 7B foundation models on four existing training corpora.
Key Capabilities
- Table-specific Instruction Following: Fine-tuned on the
tablebench_train.jsoncorpus fromdnaihao/Table-Instructs, making it adept at understanding and processing table-related instructions. - Research Tool: Serves as a valuable resource for researchers studying the impact of base models and training data on the performance of Large Language Models in table-related tasks.
- Performance Analysis: Demonstrates through its development that the choice of the base model plays a more significant role in overall performance than the specific training data used for instruction tuning.
Good for
- Academic Research: Ideal for researchers and practitioners interested in the meta-evaluation of Table LLMs and understanding the underlying factors influencing their effectiveness.
- Comparative Studies: Useful for comparing the performance of different base models when subjected to similar instruction-tuning datasets for table-based tasks.
- Understanding Model Dynamics: Provides insights into the interplay between foundational model architecture and specialized training data in achieving robust performance in specific domains like table processing.