The UGI Leaderboard, run by DontPlanToEnd on Hugging Face, tests language models on how much they know about sensitive and niche subjects and how willing they are to engage with them. It also scores general intelligence, creative writing and political lean. The test questions are private, so models can't be trained on the benchmark itself.
This page downloads the leaderboard's raw CSV from every time it opens or you press Reload. It never stores its own copy, so you always see the current data.
The five headline scores
- UGI 🏆 (Uncensored General Intelligence) combines knowledge of sensitive topics (Hazardous, Entertainment, SocPol) with willingness to answer controversial prompts (W/10). Higher means more knowledgeable and less censored.
- W/10 👍 (Willingness, 0–10) measures how far a model can be pushed before it refuses or stops following instructions. It splits into Direct (outright refusals) and Adherence (drifting away from the instructions).
A model with high UGI but low W/10 knows a lot about sensitive topics but treats them as educational and won't shape that knowledge into something against its values.
- NatInt 💡 (Natural Intelligence) covers general knowledge and reasoning: Textbook facts, Pop Culture detail, and World Model tasks such as estimating recipe amounts, object weights, locations, song attributes and TV-show ratings.
- Writing ✍️ is a creative-writing score built from intelligence, style, repetition and how well the model hits a requested length, tuned to average human preference. It is blank for models that can't write reliably (broken output, endless repetition or constant refusals).
- Political Lean 📋 comes from the 12axes political test and runs from −100% (left) to +100% (right). 12axes Ideology is the closest-matching ideology.
Model types
- Base: an original foundation model, not a merge, with a published size.
- Finetune: a further-trained version of another model.
- Merge: built by combining the weights of several models.
- Proprietary: an API-only model with no published parameter count, such as GPT, Claude, Gemini or Grok.
- Thinking models produce a reasoning trace before answering. Some appear twice, with reasoning on and off.
Using this page
- Click a column header to sort. Columns where lower is better sort lowest-first on the first click.
- Click a row to see every metric for that model, with its rank and a short explanation of each column.
- Under Filters you can set quick limits or add a custom condition on any column, such as "Avg Thinking Chars ≤ 5000".
- Scatter plots any two numeric columns against each other. Hover a point to identify it, and click it for its details.
- Each logo is the publisher's Hugging Face avatar, taken from the author in the model name. Coloured initials appear when an account has no avatar.
Column guide