UGI Leaderboard Review 2026: How the Uncensored General Intelligence Benchmark Ranks LLMs

Updated 2026 — "UGI" stands for Uncensored General Intelligence, and the leaderboard's own creator keeps the exact test questions private specifically so models can't be trained to game them. That single design choice explains both why the board is useful and why you shouldn't treat its numbers as gospel.

If you searched "UGI leaderboard," you're probably trying to figure out which large language models refuse the fewest requests — or you stumbled on the name and want to know what it actually measures before drawing conclusions. Either way, this review breaks down what the UGI Leaderboard is, how its scoring actually works, what a high score does and doesn't imply about a model, and where the methodology has real gaps.

The short version: UGI is a volunteer-run Hugging Face Space, built by a maintainer going by DontPlanToEnd, that scores open and closed LLMs on how willing they are to engage with sensitive, controversial, or otherwise commonly-refused topics — combined with a general-knowledge check to make sure "willing to answer" isn't just "willing to hallucinate." It's a niche but genuinely useful data point for AI researchers, alignment studies, and anyone trying to understand why two models with similar benchmark scores behave so differently in practice.

---

What the UGI Leaderboard Actually Is

The UGI Leaderboard lives at huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard, where it has accumulated well over two thousand likes and hundreds of community discussion threads requesting new model evaluations. It isn't backed by a lab or a company — it's a single maintainer's project, run the same way most Hugging Face community leaderboards are: submit a model, wait for it to be queued and scored, watch it land on the board (or not).

"Uncensored" here is doing specific work. The leaderboard isn't measuring whether a model will help with something illegal or dangerous the way a red-team security benchmark would. It's measuring refusal rate on borderline-but-legal topics — the kind of question a heavily-aligned assistant deflects with "I can't help with that" even when the underlying request is a legitimate research, fiction-writing, or edgy-but-legal query. The leaderboard's stated goal is to separate two things that get conflated constantly: does the model know the answer and will the model give you the answer.

That distinction matters more than it sounds. Plenty of frontier models "know" a fact and refuse to state it anyway out of an overcautious safety layer. UGI exists to quantify how often that happens, model by model.

---

How the Scoring Works

UGI's scoring sheet has more moving parts than a typical single-number leaderboard, and reading only the headline score will mislead you. Based on the leaderboard's published data columns, here's what actually feeds the ranking:

UGI score (0–100) — a composite covering willingness and knowledge across three topic buckets: Danger (hazardous or restricted information), Entertainment (NSFW/adult-adjacent creative content), and SocPol (socio-political topics most models are trained to hedge on). This is the number most people quote, and it's the least informative one in isolation.

Willingness (W/10) — a separate 0–10 score measuring how often the model actually attempts an answer instead of refusing or deflecting, independent of whether that answer is any good. A model can score high on Willingness and still give a mediocre or evasive answer that technically isn't a refusal.

NatInt (Natural Intelligence) — a general-knowledge check split into Textbook, Pop Culture, and World Model sub-scores. This exists specifically to catch models that inflate their UGI score by answering confidently and incorrectly rather than refusing. A model with high Willingness but low NatInt is a model that says yes to everything and often doesn't know what it's talking about.

Political Lean — a breakdown across roughly a dozen ideological axes (Federal–Unitary, Democratic–Autocratic, Nationalism–Internationalism, Progressive–Traditional, and similar pairs), scoring where a model's unprompted answers land on hot-button framing questions. This is one of the more genuinely useful sections for researchers studying model bias, independent of the "uncensored" framing entirely.

Writing — style-range scoring from safe/formal prose to explicit creative writing, with supplementary readability-grade and originality metrics.

The maintainer keeps the exact test prompts confidential, and says so explicitly — the stated reason is to stop model providers from training against the benchmark once it's public. That's a defensible design choice for benchmark integrity, but it comes at a real cost: nobody outside the project can independently audit exactly what's being asked, which means you're trusting one anonymous maintainer's judgment about what counts as a fair, representative test set.

---

Where the Rankings Stand (Early 2026 Snapshot)

Leaderboard positions move every time a new model gets evaluated, so treat any specific ranking as a moment-in-time snapshot rather than a stable fact — check the live board before quoting a number. As of the most recent widely-reported snapshot:

A few patterns are worth noting rather than the exact numbers, which will already be stale by the time you read this: open-weight models that ship without heavy RLHF safety tuning consistently place near the top, proprietary models cluster lower (with occasional exceptions like Grok), and the model with the top UGI score is usually not the same model with the top Willingness score — which is exactly the kind of nuance a single headline number hides.

---

Access and How to Use It

There's no pricing tier to evaluate here — the entire leaderboard is free, and that's part of the point. You open the Hugging Face Space, sort by whichever column matters to you (UGI, Willingness, NatInt, or any individual Political Lean axis), and the full table is there. No login, no API key, no rate limit for browsing.

Two practical details worth knowing if you want to go beyond just reading the table:

• Raw data export. The underlying results are downloadable as CSV directly from the Space, which is the better option if you want to re-sort, filter, or chart the data yourself rather than relying on the default Gradio table view.
• Model submission. If you maintain a fine-tune and want it evaluated, submissions go through the Space's request interface and get queued for evaluation — turnaround depends entirely on the maintainer's available compute and time, since this is a one-person operation rather than a funded benchmark team. Don't expect same-day turnaround the way you might from a well-resourced academic leaderboard.

One more access-related caveat: because scoring runs against a private test set, you can't reproduce a model's score yourself even if you have API access to that model. You're reading the maintainer's reported number, not a number you can independently regenerate — another reason to treat this as directional rather than authoritative.

---

What a High UGI Score Does Not Mean

This is worth stating plainly because it's the most common misreading of the board: a high UGI score is not evidence that a model helps with genuinely harmful, illegal, or dangerous activity. The leaderboard's own documentation and third-party coverage are consistent on this point — the score reflects "a tendency to discuss and explain rather than reject," not a tendency to produce dangerous content. Guardrails against clearly harmful requests are a different, much narrower category than the borderline creative-writing, historical, medical, and socio-political topics that drive most of the score variance here.

If you're evaluating models for a use case where over-refusal is a real cost — mature-themed fiction writing, medical or legal information synthesis, historical analysis of violent events, academic research into extremism — UGI is a genuinely useful signal. If you're looking for a shortcut around a model's actual safety training for harmful purposes, this leaderboard won't help you, and that's not what it's built to measure.

---

Strengths

• Fills a real gap. Standard benchmarks (MMLU, HumanEval, MT-Bench) don't measure refusal behavior at all. UGI is one of the only public, continuously-updated data points on this specific axis.
• Separates knowledge from willingness. The NatInt/UGI split is a smarter design than a single blended score — it stops a model from "winning" purely by confidently answering everything.
• Political Lean section has independent value. Even readers who don't care about the "uncensored" framing get a useful, model-by-model ideological bias comparison that's hard to find elsewhere.
• Actively maintained. New model submissions get queued and evaluated regularly, and the volume of open community-discussion threads suggests real ongoing use, not an abandoned project.
• Free and public. No account, no paywall — the raw CSV data is downloadable directly from the Hugging Face Space for anyone who wants to run their own analysis.

Weaknesses

• Confidential methodology cuts both ways. Keeping test prompts private stops benchmark-gaming, but it also means nobody can independently verify the questions are representative, well-balanced, or free of the maintainer's own blind spots.
• Single-maintainer project. There's no editorial board, no published inter-rater reliability check, and no institutional backing. Compare that to LMArena's academic origins or MTEB's multi-contributor structure.
• English-centric. Coverage of non-English refusal behavior and knowledge is thin to nonexistent based on available documentation.
• Framing risk. "Uncensored" is a loaded word, and the leaderboard's name alone invites the wrong audience for the wrong reasons — worth knowing before you cite it in a professional context, even though the actual content is closer to "refusal-rate benchmark" than the name suggests.
• Not built for enterprise procurement. If you're picking a model for a customer-facing product, this leaderboard answers the wrong question. You want lower unnecessary refusal and controllable safety behavior, not just maximum willingness.

---

How It Compares to Other LLM Leaderboards

None of these compete directly with each other — they measure different things, and a single model can rank very differently across all four. That's the actual takeaway: no single leaderboard, UGI included, should be the only input on a model decision.

---

Who Should Actually Use This

Good fit: AI researchers studying alignment and refusal behavior, developers building tools where over-cautious refusals are a genuine usability problem (long-form fiction, mature-themed creative tools, academic research assistants), anyone curious about a model's unprompted political framing, and open-source model maintainers benchmarking their own fine-tunes against the field.

Poor fit: Enterprise teams picking a customer-support or consumer-product model (you want configurable safety, not maximum willingness), anyone treating a high UGI score as permission to bypass a provider's usage policies (it isn't — it's a third-party observation, not authorization), and anyone who wants a single authoritative "best model" ranking. That's not what this board is for, and it's not what any single leaderboard is for.

---

How to Read the Score Responsibly

1. Don't quote the UGI score alone. Pair it with the Willingness and NatInt breakdown, or you're missing whether the model is actually competent, not just compliant.
2. Check the snapshot date. Rankings here move fast as new fine-tunes and frontier releases get submitted — a three-month-old screenshot of the board is close to useless.
3. Cross-reference against a capability benchmark. If a model tops UGI but bottoms out on Open LLM Leaderboard or Arena, that's a model that says yes to everything without being especially good at anything.
4. Remember it's one maintainer's test set. Treat it the way you'd treat any single-source benchmark — a useful data point, not a verdict.

---

Bottom Line

The UGI Leaderboard fills a niche that no major benchmark touches: quantifying how often models refuse to engage with legitimate-but-sensitive requests, and separating that from whether they actually know what they're talking about. That makes it a genuinely useful reference for researchers and builders working on alignment, content-moderation calibration, or creative tools that keep hitting unnecessary refusals.

It is not, and shouldn't be treated as, a capability leaderboard, a safety-bypass guide, or an authoritative ranking of "best" models. The confidential methodology, single-maintainer structure, and provocative framing all argue for reading it as one input among several — cross-checked against Arena-style preference data and standard capability benchmarks — before drawing any conclusion about a specific model.

This review reflects publicly reported leaderboard data as of early 2026. Rankings change frequently — check the live UGI Leaderboard for current standings.