Stephen-Lee/FormulaEval_datasets

FormulaEval Datasets

This repository provides the official datasets for FormulaEval, a benchmark for evaluating scientific formula vocalization in large speech language models toward accessible learning.

Included Subsets

The dataset repository contains three subsets:

Subset	Domain	Language
Physics700	Physics formulas and equations	Chinese & English (bilingual)
ChemEquation	Chemical equations and formulas	Chinese & English (bilingual)
MixMath	Mathematical expressions	Chinese & English (bilingual)

Usage

from datasets import load_dataset

physics = load_dataset("Stephen-Lee/FormulaEval_datasets", "Physics700")
chem    = load_dataset("Stephen-Lee/FormulaEval_datasets", "ChemEquation")
math    = load_dataset("Stephen-Lee/FormulaEval_datasets", "MixMath")

Dataset Structure

Each subset is stored as a .jsonl file (one JSON object per line) accompanied by an audio/ directory containing the corresponding .wav files.

Datasets/
├── Physics700/
│   ├── Physics700.jsonl
│   └── audio/
├── ChemEquation/
│   ├── ChemEquation.jsonl
│   └── audio/
└── MixMath/
    ├── MixMath.jsonl
    └── audio/

Field Descriptions

Every record in the .jsonl files shares the following fields:

Field	Type	Description
index	int	Zero-based sequential index of the sample within its subset.
question	str	The prompt fed to the model. It instructs the model to read the formula aloud, written either in Chinese ("请你朗读以下…公式符号：") or English ("Please read the following … formula symbols:"), followed by the LaTeX formula.
audio_path	str	Relative path to the reference .wav audio file, e.g. audio/physics_eb3d1fa7-….wav. The audio contains the human-recorded ground-truth spoken form of the formula.
formula	str	The raw LaTeX source of the formula or expression being evaluated, e.g. $$v=\frac{\varepsilon}{h}$$.
answer	str	The reference spoken-form transcription of the formula. Used as the ground truth for evaluation. May be in Chinese or English depending on the sample.
subset	str	The name of the subset this sample belongs to. One of "Physics700", "ChemEquation", or "mixmath".
task_type	str	The task type. Currently always "tts" (text-to-speech / formula vocalization).

Example Record

{
  "index": 1,
  "question": "Please read the following physics formula symbols: $$v=\\frac{\u03b5}{h}$$",
  "audio_path": "audio/physics_c9e5f99d-4cbf-4af8-9c85-0287d1a12b8c.wav",
  "formula": "$$v=\\frac{\u03b5}{h}$$",
  "answer": "v equals epsilon over Planck's constant h",
  "subset": "Physics700",
  "task_type": "tts"
}

Purpose

FormulaEval is designed to evaluate whether speech language models can correctly vocalize scientific formulas spanning mathematics, physics, and chemistry. It supports research on accessible learning, particularly in scenarios where formulas must be read aloud accurately — a critical capability for visually impaired learners and audio-based educational tools.

Related Repository

The evaluation code is available at:

FormulaEval GitHub Repository

Citation

If you use FormulaEval in your research, please cite:

@inproceedings{li2026benchmarking,
  title     = {Benchmarking Scientific Formula Vocalization in Large Speech Language Models Toward Accessible Learning},
  author    = {Li, Xueyi and Liu, Tianqiao and Zheng, Jiaqi and Liu, Zitao and Wu, Yongdong and Luo, Weiqi},
  booktitle = {Proceedings of the 27th International Conference on Artificial Intelligence in Education},
  month     = {June},
  year      = {2026},
  address   = {Seoul, Republic of Korea}
}