I thought `surely I could get a large language model to make a more calm alternate text while honouring the original intent as best as possible?` Well it works sorta okay, I've got a slow and costly implementation using my own Mistral credits, that I am paying for (free credits are welcome!)
I've just set it up to collect data for fine-tuning a translator (encoder-decoder) model later, I've also identified some datasets I could use, as well as collecting from people using the Space. I have a roadmap on the GitHub page.
I tried using Mixtral with llama.cpp's Python interface on RunPod Severless and HuggingFace (with JSON grammars enforced), but it wasn't reliable enough so I am just using the Mistral API.
Wouldn't it be great to have rose tinted glasses that show a calmer internet? Or in the moment offer people better alternatives instead of posting something inflammatory?
ChillTranslator is a tool aiming to transform toxic online comments into more constructive dialogues.
It's MIT licensed Python code with a Gradio frontend and module that might be reused elsewhere. Ideally people could just pip install it and have it just work right away but I fear there will be rate limits and moderation limits hit easily, so I figure a small translator model is a good solution. Currently I know it's going to be biased in the way that LLM's usually are, and it might not work very well, and it's buggy, but it's a start! There is a demo here: https://huggingface.co/lukestanley/ChillTranslator
What do you think?