|
|
|
|
|
|
by probably_wrong
1136 days ago
|
|
|
> Until uncensored models are generally available, these novelty models will always be less-than. The most popular generative model on HuggingFace at the time of this comment is Pygmalion 6b [1], a model that I believe is fine tuned on top of Alpaca to generate porn. I couldn't find the data source, though, so I don't know on what kind. And Facebook's "leaked" LLaMa, while not fine-tuned for conversation, has several warnings on its potential for offensive content. If I read the instructions correctly, mlc-ai is loading "plain" Alpaca which is great for conversation but, as you notice, rather conservative. I don't think this is a bad idea - perhaps it's better if we don't inflict racist AI on unsuspecting users. Try shopping around for other models. Edit: I repeated your experiment with other models (but another library). They had no objections against generating offensive-yet-unfunny jokes. [1] https://huggingface.co/PygmalionAI/pygmalion-6b |
|
|
With MLC being the first LLM-in-a-box to run on my M2 at faster than a token per minute, I'm impressed at the speed but also disappointed at the quality of the experience. For those interested in the outcome, it failed all 3 tests, which is not unexpected for a small model like this.
Using/producing models with censorship included voluntarily demonstrates a willingness to hobble the technology for peripheral reasons that do not directly correlate with the advancement of the field. For that reason, this is a disqualifying characteristic in the capacity of my own use on the basis that social sensibilities and decency varies across cultural and regional lines, anything so trivial as a crass joke being limited is such a low bar that other things of much more grave concern will undoubtedly be tampered with or limited, and not always in ways the authors intended.
Self-hindering behavior will not be the positive we think it will be, as with most measures to correct injustices with data.