dnaihao/phi-3-mini-tablellama
dnaihao/phi-3-mini-tablellama is a 4 billion parameter instruction-tuned causal language model, replicated from Microsoft's Phi-3-mini-4k-instruct. It is specifically fine-tuned on the TableLlama corpus for tasks involving tabular data. This model extends the research presented in the EACL 2026 Findings paper, focusing on the effects of model and data on Table LLMs, and is optimized for understanding and generating responses based on structured table information.
Loading preview...
Overview
dnaihao/phi-3-mini-tablellama is a 4 billion parameter instruction-tuned model, replicated from Microsoft's Phi-3-mini-4k-instruct. It has been specifically fine-tuned using the tablellama_train.json corpus from the dnaihao/Table-Instructs dataset. This model serves as an additional artifact for the EACL 2026 Findings paper "What Really Matters for Table LLMs? A Meta-Evaluation of Model and Data Effects", extending its experiments by adding another base-model variant trained on the same specialized corpus.
Key Capabilities
- Table-specific Instruction Following: Optimized for understanding and generating responses based on structured tabular data.
- Research Replication: Provides a fine-tuned variant of Phi-3-mini for meta-evaluation of Table LLMs, complementing the main research grid.
- Open-source Training: Trained using LLaMA-Factory with a learning rate of 5e-7, allowing for transparency and reproducibility.
Good for
- Research on Table LLMs: Ideal for researchers studying the impact of base models and training data on performance with tabular data.
- Applications requiring Table Understanding: Suitable for tasks that involve extracting information from, querying, or generating content based on tables.
- Extending EACL 2026 Research: Directly supports and extends the experimental findings of the associated paper.