AI / interpretability
Political Alignment
Tools for examining political alignment in language models.
experimental
Political Alignment evaluates language models on the political compass and analyses their outputs for pairs of concepts.
The numbers shown are README-reported Euclidean distances between mean-pooled output logits. They are not political scores and were not reproduced here.
AI / interpretability
“marx” ↔ “stalin”
- GPT-2 base
- 1,985
- MarxGPT-2 aligned
- 2,249
README-reported Euclidean distances between mean-pooled output logits, not political scores.