github_com/Sumandora/remove-refusals-with-transformers

This is a crude, proof-of-concept implementation to remove refusals from an LLM model without using TransformerLens. This means, that this supports every model that HF Transformers supports*.

The code was tested on a RTX 2060 6GB, thus mostly <3B models have been tested, but the code has been tested to work with bigger models as well.

*While most models are compatible, some models are not. Mainly because of custom model implementations. Some Qwen implementations for example don't work. Because model.model.layers can't be used for getting layers. They call the variables so that, model.transformer.h must be used, if I'm not mistaken.
--
修改:tgfbeta FROM 111.162.221.*
FROM 111.162.221.*