Arena Human Preference 140K: Text-Only LLM Battle Votes

140,000 user votes comparing two language models on a conversation, collected via the LM Arena platform. Each row contains a single vote with the full conversation history and metadata like the winning model and evaluation session. The dataset was created by lmarena-ai and last updated in August 2025.

Use Cases

• Training reward models for reinforcement learning from human feedback (RLHF) based on pairwise comparison votes.
• Analyzing model performance and user preferences across different conversation contexts using the provided conversation history.
• Benchmarking language models in a head-to-head 'battle' format based on the winner field.
• Studying evaluation consistency across sessions using the evaluation_session_id and evaluation_order metadata.

Strengths

• Contains 140,000 individual human preference judgments, providing a substantial signal for model training.
• Includes full conversation history for each vote, allowing for contextual analysis of model outputs.
• Metadata such as evaluation_session_id and evaluation_order supports analysis of voting patterns.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

lmarena-ai via the Hugging Face platform.

Collection Method

User votes collected in the text-only category of an evaluation arena.

Freshness

Last updated 2025-08-01 05:03:39; freshness should be verified.