Snowflake/dare-bench

---
language:
- en
arxiv: TODO
task_categories:
- text-generation
- question-answering
tags:
- agent
- data-science
- tool-use
- reinforcement-learning
- sft
pretty_name: DARE-Bench
license: apache-2.0
configs:
- config_name: default
data_files:
- split: train
path: "train_viewer.jsonl"
- split: eval
path: "eval_viewer.jsonl"
---

# DARE-Bench

**[ICLR 2026]** DARE-Bench: Evaluating Modeling and Instruction Fidelity of LLMs in Data Science

Fan Shu¹, Yite Wang², Ruofan Wu¹, Boyi Liu², Zhewei Yao², Yuxiong He², Feng Yan¹

¹University of Houston   ²Snowflake AI Research

## 🔎 Overview

**DARE-Bench** (ICLR 2026) is a benchmark for evaluating LLM agents on data science tasks, focusing on modeling and instruction fidelity.
This Hugging Face repository provides a selected subset of the full benchmark for public release.

### ✨ Highlights

- ✅ Selected subset of the full benchmark
- 📚 Splits: Train **2,137** • Eval **162**
- 🗃️ Database assets distributed as zip files for convenient download

## 📦 Dataset Contents

| Split | Files | #Entries |
| --- | --- | ---: |
| Train | `train/question_list.json`, `train/databases.zip` | 2,137 |
| Eval | `eval/question_list.json`, `eval/databases.zip` | 162 |
| SFT | `sft_data/` | — |

### 🗂️ File layout

- `sft_data/`: supervised fine-tuning trajectories.
- `train/question_list.json`: training tasks (JSON array).
- `train/databases.zip`: database assets for training tasks.
- `eval/question_list.json`: evaluation tasks (JSON array).
- `eval/databases.zip`: database assets for evaluation tasks.

## 🔗 Resources

Related resources are also available, please check:

| Resource | Link |
| --- | --- |
| 📄 Paper | [arxiv.org/abs/2602.24288](https://arxiv.org/abs/2602.24288) |
| 💻 Code | [Snowflake-Labs/dare-bench](https://github.com/Snowflake-Labs/dare-bench) |

## 📜 License

- The `databases/` assets are subject to their **original source licenses**.
- All other files in this repository are released under **Apache-2.0**.

## 📝 Citation

```bibtex
@inproceedings{shu2026darebench,
title={DARE-Bench: Evaluating Modeling and Instruction Fidelity of LLMs in Data Science},
author={Shu, Fan and Wang, Yite and Wu, Ruofan and Liu, Boyi and Yao, Zhewei and He, Yuxiong and Yan, Feng},
booktitle={International Conference on Learning Representations},
year={2026}
}
```