DARE-Bench: 1,000+ Data Science Tasks for LLM Agent Evaluation

DARE-Bench contains between 1,000 and 10,000 records designed to evaluate Large Language Model (LLM) agents on data science modeling and instruction fidelity. Developed by Snowflake AI Research and the University of Houston for ICLR 2026, the dataset focuses on tool-use and text generation within data science workflows. It provides a selected subset of a larger benchmark for testing how models handle complex data manipulation instructions.

Use Cases

• Evaluating LLM agent fidelity using Pandas and Polars libraries
• Benchmarking text generation for data science question answering
• Supervised Fine-Tuning (SFT) for data science tool-use agents

Strengths

• 1,000 to 10,000 records
• ICLR 2026 peer-reviewed research origin
• Supports Pandas and Polars library integration for tool-use evaluation

Limitations

• Subset of full benchmark
• Specific column schema and field definitions are not documented in the metadata

Provenance

Source

Snowflake AI Research and University of Houston

Freshness

Last updated March 2026.