Spreadsheet Arena: 555 Pairwise Human Preference Votes on LLM-Generated Spreadsheets

Longitude-Labs released this dataset of 555 pairwise human preference votes over LLM-generated spreadsheets. The data spans 124 distinct user-submitted prompts and 17 models, accompanying the Spreadsheet Arena paper. It was last updated on May 27, 2026.

Use Cases

• Benchmarking LLM performance on spreadsheet generation tasks based on 124 distinct user prompts.
• Training or fine-tuning reward models for spreadsheet generation based on 555 pairwise human preference votes.
• Analyzing model preferences and biases in structured data output across 17 different models.

Strengths

• Contains 555 pairwise human preference votes, providing a basis for preference modeling.
• Covers 124 distinct user-submitted prompts, offering a variety of test scenarios.
• Evaluates outputs from 17 different models, allowing for comparative analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the main battles.csv file is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the specific user-submitted prompts and voting methodology.

Provenance

Source

Longitude-Labs

Collection Method

Likely collected via a platform for human evaluation of LLM-generated spreadsheet outputs.

Freshness

Last updated 2026-05-27 20:42:25; freshness should be verified.