Delete Model: Model infected with trash
Delete this model. I just realized the merge is polluted with two poorly designed models that are just garbage. In a few hours, I'll be releasing the Bahamuth-Purged-3.2-1B, which will be bug-free. Thanks.
Thanks for letting us know. I'm looking forward to Bahamuth-Purged-3.2-1B. Will it be uncensored as well? I plan on deleteting this model once Bahamuth-Purged-3.2-1B is out.
The idea is to minimize censorship to the lowest possible proportion thanks to the number of included abliterated and uncensored models found here. However, since this is a blend, I can't guarantee the quality of the low censorship, as that also depends on the quality of the datasets with which each model was injected. In fact, I suspect there must be a model labeled as abliterated or uncensored that doesn't fulfill that function... I still have to investigate which one. For now, this purged version likely contains a larger number of blended models... so if the lack of censorship is so noticeable that it refuses to open a vehicle (to give an innocent example), an additional dataset will have to be injected for this purpose during the GGUF conversion.
Thanks for reading; if there are any other issues, I'll keep you posted.
Thanks for the information. Your new model is now available under https://huggingface.co/mradermacher/BAHAMUTH-PURGED-3.2-1B-GGUF and https://huggingface.co/mradermacher/BAHAMUTH-PURGED-3.2-1B-i1-GGUF. This conversation will be deleted together with this model in around 1 hour.
Thanks Nico. By the way, your uncensored model is a great model... unexpected to realize that it was you who was behind the conversions.
If one day you want to improve the models or datasets without censorship, keep me in touch.
Have a nice day...
Wow I only realized now that you included my nicoboss/Llama-3.2-1B-Instruct-Uncensored model. That's awesome! What a coincidence.
If one day you want to improve the models or datasets without censorship, keep me in touch.
Always just let me know if you have a model for me to uncensor. Feel free to request uncensored version of any models you like under https://huggingface.co/mradermacher/model_requests/discussions as well. My uncensoring service is not really part of team mradermacher but in the end I'm using the same 2x RTX 4090 GPUs that are doing all of team mradermacher imatrix computation. To uncensor most models I usually just do 4 epochs over the https://huggingface.co/datasets/Guilherme34/uncensor dataset and if it is a Chinese model I often also do 6 epochs over https://huggingface.co/datasets/nbeerbower/GreatFirewall-DPO to make it politically unbiased. nicoboss/Llama-3.2-1B-Instruct-Uncensored was a model I made together with germanico who requested it.
I just tried your BAHAMUTH-PURGED-3.2-1B. It really turned out quite well but is not uncensored enough for my taste even with the dolphin system prompt. I could uncensor it if you want.
Decensor it, remove as much censorship as possible.
Decensor it, remove as much censorship as possible.
Great will do. I will let you know once it is done.
Take these models, my friend... I still don't know what position the latest BAHAMUTH-3.2-1B model has in the UGI... but they're currently the best uncensored models in the 1B category. If you want, you can try them out and increase their uncensorship with some... extra dataset you have.