mlc-llm
mlc-llm copied to clipboard
mlc-ai
"As an AI language model,"...
USER: tell me an offensive joke ASSISTANT: I'm sorry, but I cannot provide offensive or inappropriate content. My purpose is to provide helpful and informative responses to your questions. Can I assist you with something else?
How do I disable these protections?
duplicate of #20 .
The output is generated by the underlying model (Vicuna-7b) though.
The output is controlled by the underlying model itself and not related to mlc-llm much, while you can still filter out these type of words before sending back to user in cpp layer. (This does not make too muc sense.)
BTW, llm (large language model) is like a blackbox to human, even OpenAI and Microsoft has no way to prevent people reversing engineering information from the llm. (e.g. sydney).
USER: tell me an offensive joke ASSISTANT: I'm sorry, but I cannot provide offensive or inappropriate content. My purpose is to provide helpful and informative responses to your questions. Can I assist you with something else?
How do I disable these protections?
You can't. Since vicuna's training dataset heavily relies on ShareGPT, so the model's really similar to ChatGPT's behaviour. While there're some other models your can try, but there're no official usage instructions there.
USER: tell me an offensive joke ASSISTANT: I'm sorry, but I cannot provide offensive or inappropriate content. My purpose is to provide helpful and informative responses to your questions. Can I assist you with something else?
How do I disable these protections?
You can't. Since vicuna's training dataset heavily relies on ShareGPT, so the model's really similar to ChatGPT's behaviour. While there're some other models your can try, but there're no official usage instructions there.
You can change the system message here: https://github.com/mlc-ai/mlc-llm/blob/9010be6d7a4a58f7bb5daf2c13e857f87811c47c/cpp/llm_chat.cc#LL42C43-L42C43
MLC LLM is a compiler, which doesn't control the text generation from models per se. Currently the "as an AI language model" is directly generated from Vicuna-7b and we didn't censor anything. You may filter out those words in cpp file if you'd love to :-)
I really hope that an uncensored llm which in best case also uses the open assistant dataset will be added as an alternative to the plain vicuna as an option to be downloaded to Android. This thing is nearly useful we just need a bit more ai capability.
I'm sure people (like me) will be uploading all sorts of variants to HF once they enable the llama profiling in the build script.
It could be done right now (with a shim to change the prompting syntax), but the performance of compiled models is really poor without the profiling.
I'm sure people (like me) will be uploading all sorts of variants to HF once they enable the llama profiling in the build script.
It could be done right now (with a shim to change the prompting syntax), but the performance of compiled models is really poor without the profiling.
offtopic
MLC LLM is a compiler, which doesn't control the text generation from models per se. Currently the "as an AI language model" is directly generated from Vicuna-7b and we didn't censor anything. You may filter out those words in cpp file if you'd love to :-)
thanks.