TAT-LLM: A Specialized Language Model for Discrete Reasoning over Tabular and Textual Data

Paper: https://arxiv.org/abs/2401.13223

Code: https://github.com/fengbinzhu/TAT-LLM

Introduction

We present TAT-LLM, a specialized language model crafted through the innovative Step-wise Pipeline approach, focusing on the nuanced realm of tabular and textual question answering (QA). This model is the fruit of rigorously fine-tuning the LLaMA 2 architecture with a novel dataset, autonomously generated from expertly annotated resources. TAT-LLM stands at the intersection of tabular comprehension and textual analysis, engineered to excel by embodying three fundamental phases: Extraction, Reasoning, and Execution. Our empirical findings illuminate TAT-LLM's remarkable capability to eclipse traditional benchmarks, surmounting even the most advanced models and colossal language models such as GPT-4 across a suite of demanding financial QA tasks like FinQA, TAT-QA, and TAT-DQA. This endeavor not only sets a new standard for task-specific language models but also paves the way for future explorations in optimizing smaller models for highly specialized functions.

Training

We train our TAT-LLM model in various sizes, including 7B, 13B, and 70B, using different methods such as parameter-efficient fine-tuning and full-parameter fine-tuning of LLaMA 2 on a combination of financial data from the FinQA, TAT-QA, and TAT-DQA training sets(🤗HuggingFace Repo). To refine accuracy, we introduce an External Executor, enhancing the model by processing intermediate outputs to derive conclusive answers. Please refer to the paper for more details.

Inference & Evaluation

Please refer to code here

Citation

If you find this model helpful, please consider citing our paper:

@misc{zhu2024tatllm,
      title={TAT-LLM: A Specialized Language Model for Discrete Reasoning over Tabular and Textual Data},
      author={Fengbin Zhu and Ziyang Liu and Fuli Feng and Chao Wang and Moxin Li and Tat-Seng Chua},
      year={2024},
      eprint={2401.13223},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}