dnaihao/mistral-v0.3-tablellama
dnaihao/mistral-v0.3-tablellama is a 7 billion parameter instruction-tuned language model, replicated from the TableLlama architecture and built upon Mistral-7B-Instruct-v0.3. It is specifically fine-tuned for table-related tasks using the TableLlama instruction-tuning corpus. This model is part of a research effort to evaluate the impact of base models versus training data on table LLM performance, demonstrating that the base model choice is a more dominant factor. Its primary use case is for tasks involving structured data in tables.
Loading preview...
Overview
dnaihao/mistral-v0.3-tablellama is a 7 billion parameter instruction-tuned model, a replication of the TableLlama architecture, built on the Mistral-7B-Instruct-v0.3 foundation. It was fine-tuned using the tablellama_train.json corpus from the dnaihao/Table-Instructs dataset. This model was developed as part of the EACL 2026 Findings paper, "What Really Matters for Table LLMs? A Meta-Evaluation of Model and Data Effects," which investigates the contributions of base models versus training data in table-focused LLMs. The research indicates that the choice of the base model significantly outweighs the impact of the training data itself.
Key Capabilities
- Table-specific instruction following: Optimized for understanding and generating responses related to tabular data.
- Research replication: Provides a publicly available instance of the TableLlama model for further research and evaluation.
- Foundation for meta-evaluation: Serves as a component in a broader study comparing different base models and training corpora for table tasks.
Good for
- Research on table-based LLMs: Ideal for academics and researchers studying the effectiveness of various base models and instruction-tuning datasets for tabular data processing.
- Developing table-aware applications: Can be used as a base for applications requiring an understanding of structured data within tables.
- Benchmarking: Useful for evaluating and comparing the performance of different LLMs on table-related benchmarks, especially in the context of the associated research paper.