Data sources
LMArena
Benchmark values come from the public lmarena-ai/leaderboard-dataset on Hugging Face. Tokenlane displays published values for explicitly mapped model variants.
Method and interpretation
In comparison arenas, people evaluate anonymized model responses side by side. LMArena derives Bradley-Terry ratings from these preferences. Often called Arena Elo, these values are relative ratings within an arena and category, not percentages.
Confidence intervals express uncertainty in the published rating. Small rank differences are not a reliable quality gap when intervals overlap. Vote counts, ranks and totals are shown with the publication date. Results do not replace evaluation for a particular use case.
Agent Arena uses separately labeled IPS estimates with observation and, where available, session counts. These values can be negative and must not be equated with Bradley-Terry ratings or vote counts from other arenas.
Mapping and updates
Variants with and without reasoning stay separate. All imported categories are accessible through the arena selector and category table. The catalog sorts by the selected arena rating of the first confirmed variant and shows its published rank. Agent IPS estimates appear only on model pages. An ambiguous or missing value remains empty. Models without a mapping or published measurement have no benchmark section.
The dataset is fetched daily on the server. If retrieval fails, the previous snapshot remains visible with its publication date. The dataset license is checked before every import; a license other than CC BY 4.0 stops the import. Values are not converted or combined into a separate overall score.
We show published ranks, excerpts and mapped model variants. Labels and number formats are adapted for Tokenlane; measurement values remain unchanged.