From 04f284d202caf7b27c1a170cb5d7b1013dddb396 Mon Sep 17 00:00:00 2001 From: Ettore Di Giacinto Date: Mon, 7 Oct 2024 09:56:33 +0200 Subject: [PATCH] models(gallery): add gemma-2-9b-it-abliterated (#3743) Signed-off-by: Ettore Di Giacinto --- gallery/index.yaml | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+) diff --git a/gallery/index.yaml b/gallery/index.yaml index 22978dd28787..42f375b3c440 100644 --- a/gallery/index.yaml +++ b/gallery/index.yaml @@ -2383,6 +2383,24 @@ - filename: Gemma-2-2B-ArliAI-RPMax-v1.1-Q4_K_M.gguf sha256: 89fe35345754d7e9de8d0c0d5bf35b2be9b12a09811b365b712b8b27112f7712 uri: huggingface://bartowski/Gemma-2-2B-ArliAI-RPMax-v1.1-GGUF/Gemma-2-2B-ArliAI-RPMax-v1.1-Q4_K_M.gguf +- !!merge <<: *gemma + name: "gemma-2-9b-it-abliterated" + urls: + - https://huggingface.co/IlyaGusev/gemma-2-9b-it-abliterated + - https://huggingface.co/bartowski/gemma-2-9b-it-abliterated-GGUF + description: | + Abliterated version of google/gemma-2-9b-it. + + The abliteration script (link) is based on code from the blog post and heavily uses TransformerLens. The only major difference from the code used for Llama is scaling the embedding layer back. + + Orthogonalization did not produce the same results as regular interventions since there are RMSNorm layers before merging activations into the residual stream. However, the final model still seems to be uncensored. + overrides: + parameters: + model: gemma-2-9b-it-abliterated-Q4_K_M.gguf + files: + - filename: gemma-2-9b-it-abliterated-Q4_K_M.gguf + sha256: 88d84ac9796732c10f6c58e0feb4db8e04c05d74bdb7047a5e37906a589896e1 + uri: huggingface://bartowski/gemma-2-9b-it-abliterated-GGUF/gemma-2-9b-it-abliterated-Q4_K_M.gguf - &llama3 url: "github:mudler/LocalAI/gallery/llama3-instruct.yaml@master" icon: https://cdn-uploads.huggingface.co/production/uploads/642cc1c253e76b4c2286c58e/aJJxKus1wP5N-euvHEUq7.png