lmarena-ai/leaderboard-dataset

Arena Leaderboard Dataset

Historical snapshots of the Arena leaderboard.

Usage

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_bash_reliability, 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.
• 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.
• On October 12, 2026, the Agent Arena bash signal was switched from bash recovery to bash reliability.

For a complete record of all leaderboard methodology changes, see the 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 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)