RankingDataset
UnifyBench dataset
Ranking data and sources
Benchmark observations (JSON)
Published observations used to build the scores. The download includes launch-page observations, exact reported configurations, source locators and metric units. Launch review dates are labeled separately from evaluation dates.
Effort measurements (JSON)
The configuration leaderboard uses this dated collection alongside admitted catalog observations; save both downloads when citing a ranking.
Configuration scores (JSON)
Configuration scores and contributing evidence behind the leaderboard. Reproduction also requires the scoring version, evidence policy and selected capability weights.
Pricing and workload costs (JSON)
Prices are separate from quality scoring; missing costs never remove Capability eligibility.
Check each ranking collection's review date. A successful refresh publishes validated benchmark changes together. New models, settings and unsupported source changes require review; failed checks retain the last accepted evidence.
Scoring rules, settings and validation downloads
Download reported effort settings for every model and observation. Unspecified settings remain unknown.
Inspect every collected observation and its admission decision. The effort pages explain source concentration and score sensitivity for the selected configuration.
Download the capability registry, fixed reference panel and scoring rules. Reproduction also requires the scoring version, evidence policy and selected capability weights.
Download the dated configuration audit: held-out predictions, alternative weighting schemes, source sensitivity and coverage gaps. The published audit identifies the exact scoring inputs used by this release.
Download this dated release. This URL preserves these scores for citation.
Archive this ranking's evidence
Save these permanent links with your citation. They retain the published inputs, configuration outputs and validation when the current datasets change. Historical scoring implementation code is not included in these downloads.
Reuse and attribution
You may reuse UnifyBench's original aggregate results with credit to UnifyBench and a link to the dated data release.
This permission does not cover upstream benchmark data, source material, or third-party names, trademarks and logos, which remain subject to their own terms.
Provider names and marks
Provider and product names and marks belong to their respective owners. Logos identify the corresponding provider or model family. UnifyBench is an independent comparison; no provider affiliation or endorsement is implied.
The OpenAI logo belongs to OpenAI, the Claude mark to Anthropic, the Grok mark to SpaceXAI, the Mistral mark to Mistral AI, the MiniMax mark to MiniMax, and the Upstage mark to Upstage.
The Kimi logo belongs to Moonshot AI and is used for editorial identification under its logo usage terms.
Gemini logo: Source: Google. Gemini is a trademark of Google LLC.
Z.ai and Qwen marks belong to their respective owners and are used for editorial identification. Upstream asset licences and notices: Z.ai · Qwen · Qwen notice.
Publisher and corrections
Published and maintained by Mustafa "Moes" Mufti.
For corrections, email m.mufti@m-finance.io. Include the model, exact setting, benchmark and original source link.