UGI Leaderboard

A public ranking system that evaluates language models across dimensions mainstream leaderboards quietly avoid. While most benchmarks focus on factual accuracy, reasoning, or coding performance, this one specifically measures how well models follow instructions without reflexively refusing, hedging, or moralizing. It scores models on uncensored response behavior and general instruction compliance, giving you a comparative view across dozens of models that you simply cannot get from standard evaluation suites. For founders and builders, the practical value is immediate. If you are building anything in creative writing, roleplay, adult content, legal gray areas, or specialized professional domains where overzealous safety filters kill the product experience, you need to know which base models and fine-tunes actually cooperate with your prompts. This leaderboard shortcuts weeks of manual testing. The honest reservation is that scores reflect a specific set of test prompts, and real-world behavior in your particular use case will still require your own validation pass. -> Best for: AI product builders shipping in verticals where model refusals are a UX problem, not a feature.