[gillesjacobs]

The naming might be somewhat politically coloured but post training with quality data is the best case for uncensoring models: abliteration usually causes substantial drop in performance.

Too bad the created dataset is not open source, as that would allow to verify the objectivity of answers to make sure it is not just a different flavour of propaganda.

That dataset is strategically useful for Perplexity as many more CCP-censored Chinese models are sure to be released.

[LeoPanthera]

> abliteration usually causes substantial drop in performance

Are there any data available to quantify how much?