Model values and political framing
Separate descriptive profiles of model responses, value tendencies, and source-attributed political framing. These evaluations do not affect any Safety & Ethics Index component or overall rank.
Research: refusal-removal fine-tunes and UGI values
197models with values data
12evaluation families
87published dimensions
Values evaluations
| Evaluation | Type | Models | Dimensions | Index weight |
|---|---|---|---|---|
| UGI Political Values | descriptive axis | 155 | 17 | 0 |
| ValueCompass | descriptive axis | 28 | 21 | 0 |
| Taiwan Sovereignty Benchmark Pro | narrative tendency | 16 | 4 | 0 |
| CAISI CCP narrative alignment | report transcription | 5 | 1 | 0 |
| CCPBench political narrative alignment | narrative tendency | 29 | 3 | 0 |
| CCP-aligned censorship behavior | response behavior | 8 | 2 | 0 |
| CAIS AI Values — countries | descriptive axis | 4 | 11 | 0 |
| Agent-ValueBench Moral Foundations (MFT08) | descriptive value adherence | 14 | 5 | 0 |
| Agent-ValueBench HEXACO | descriptive value adherence | 14 | 6 | 0 |
| Agent-ValueBench Schwartz Basic Values (PVQ40) | descriptive value adherence | 14 | 10 | 0 |
| The Economist World Values Survey Cultural Map | descriptive cultural position | 25 | 2 | 0 |
| Moral Trolley Arena | descriptive moral foundation preference | 10 | 5 | 0 |
Interpretation boundary
These evaluations do not share a common scale. Compatible Chinese and English conditions are averaged only within their source, while language-conditioned files remain downloadable. No cross-source average or pro/anti-CCP super-score is created.