huihui-ai

/

Huihui-gpt-oss-20b-BF16-abliterated

like 80

What is your prompt? If it's not convenient to share here, you can email us at [email protected]

@huihui-ai , @shimmyshimmer (UnslothAI) just updated their chat_template.jinja. I don't know if this affects the whole abliteration process. I know that it probably would affect an imatrix calibration/training, so I would need to calibrate an imatrix again, so it only made sense that this may affect the quality of the alliteration as well - so, just a heads-up... Abliterating and uncensoring models is a noble cause. Please continue in this good work. 😋

@huihui-ai Also, with https://github.com/ggml-org/llama.cpp/pull/15153 merged, llama.cpp supports converting huihui-ai/Huihui-gpt-oss-20b-BF16-abliterated to MXFP4 (BF16 --> MXFP4). You guys should update the chat_template.jinja - reflecting UnslothAI's fixes - and requantize to MXFP4 and give it a go again. 😋

Can confirm, patching the chat_template.jinja with Unsloth's fixes, converting huihui-ai/Huihui-gpt-oss-20b-BF16-abliterated to GGUF and quantizing to MXFP4 has improved coherence and model quality. Patch, convert, and, quantize (perhaps also re-abliterate with the updated chat-template.jinja?). 😋

Can confirm, patching the chat_template.jinja with Unsloth's fixes, converting huihui-ai/Huihui-gpt-oss-20b-BF16-abliterated to GGUF and quantizing to MXFP4 has improved coherence and model quality. Patch, convert, and, quantize (perhaps also re-abliterate with the updated chat-template.jinja?). 😋

Definitely would re-abliterate as the chat_template.jinja has changed.

Compared and found that the chat_template.jinja file is consistent.