leaderboard-dataset

# Arena Leaderboard Dataset Historical snapshots of the [Arena](https://arena.ai/leaderboard) leaderboard. ## Usage ```python from datasets import load_dataset # Load all historical text style control data ds = load_dataset("lmarena-ai/leaderboard-dataset", "text_style_control", split="full") # Load the current text style control leaderboard ds = load_dataset("lmarena-ai/leaderboard-dataset", "text_style_control", split="latest") # Filter to overall category ds = load_dataset("lmarena-ai/leaderboard-dataset", "text_style_control", split="latest", filters=[("category", "==", "overall")] ) # Track a specific model's rating over time ds = load_dataset("lmarena-ai/leaderboard-dataset", "text_style_control", split="full", filters=[("category", "==", "overall"), ("model_name", "==", "gpt-4o-2024-05-13")], columns=["model_name", "rating", "rank", "leaderboard_publish_date"] ) # Load raw (non-style-controlled) text ratings ds = load_dataset("lmarena-ai/leaderboard-dataset", "text", split="full") # Load the current Agent Arena leaderboard ds = load_dataset("lmarena-ai/leaderboard-dataset", "agent", split="latest") ``` ## Subsets Each arena is a separate subset, arenas with style control have an additional subset with a `_style_control` suffix . - `text`, `text_style_control` - `vision`, `vision_style_control` - `search`, `search_style_control` - `document`, `document_style_control` - `webdev` (Code Arena) - `text_to_image` - `image_edit` - `text_to_video` - `image_to_video` - `video_edit` - `agent` (Agent Arena aggregate), plus per-signal subsets: `agent_bash_recovery_steps`, `agent_praise_complaint`, `agent_steerability`, `agent_task_outcome_explicit`, `agent_tool_hallucination` ## Splits - `full`: All historically published leaderboards - `latest`: Only the most recently published leaderboards ## Notes - On January 9, 2024, the rating system was updated from Elo to [Bradley-Terry](https://en.wikipedia.org/wiki/Bradley%E2%80%93Terry_model). - On May 16, 2025, style control was made the default for text and vision arenas and the offset was adjusted to put the style control leaderboard on the same rating scale as non style control. - On July 23, 2025, frequency-based re-weighting was implemented. - Search and Webdev Arena data starts from their releases on the arena.ai (then lmarena.ai) on August 7 and November 12 2025 respectively. - Agent Arena data starts from its release on June 4, 2026. For a complete record of all leaderboard methodology changes, see the [Leaderboard Changelog](https://arena.ai/blog/leaderboard-changelog/). ## Schema | Column | Type | Description | |--------|------|-------------| | model_name | string | Model identifier | | organization | string | Model creator/organization | | license | string | Model license | | rating | float | Arena Score | | rating_lower | float | Lower confidence bound | | rating_upper | float | Upper confidence bound | | variance | float | Rating variance | | vote_count | int | Number of battles for this model | | rank | int | Rank within this leaderboard | | category | string | Leaderboard category (e.g., overall, coding, math) | | leaderboard_publish_date | string | Date that this score was published (`YYYY-MM-DD`) | Agent Arena subsets use [IPS scores](https://arena.ai/blog/agent-arena-methodology/) instead of Bradley-Terry: | Column | Type | Description | |--------|------|-------------| | model_name | string | Model identifier | | organization | string | Model creator/organization | | license | string | Model license | | score | float | IPS score (τ̂) | | score_ci_lower | float | Lower 95% confidence bound | | score_ci_upper | float | Upper 95% confidence bound | | observation_count | int | Signal observations for this model | | session_count | int | Distinct sessions (aggregate subset only) | | rank | int | Rank within this leaderboard | | category | string | Leaderboard category | | leaderboard_publish_date | string | Date that this score was published (`YYYY-MM-DD`) |