Things I’m Learning While Training SuperHOT
I have been working on SuperHOT for some time now. It is a fiction-focused finetune of LLaMa with extra focus towards NSFW outputs while also being capable of general use instructions. The main reason I’m making the model is because it is fun and serves as a good way to learn the inner workings of Transformers, dataset creation and techniques, and probing the capabilities of LLMs. There are also a lot of people who want NSFW-capable models, and they provide useful, honest feedback, especially when they don’t get what they want. Besides, it’s a fun model to use.
I’m making this page to share some of my findings in the hope that others might find it useful. I will update this page as time goes on with any information that might be important and whenever I have the time.
Background
Originally, I was working on a Langchain extension for oobabooga’s Text-Generation-WebUI. It was not a very fun project. Gradio is not fun to work with when it comes to stateful UI updates, or even encapsulation of UI and logic. I ended up hand-rolling my own UI state management system just to give a nice user experience when modifying templates, chains, etc. After some time of using Langchain, I realized I was only using a subset of the features (because they were the most useful to me), and even those features could be replicated outside of the framework very easily and save me from the unnecessary bloat. I was also displeased with the quality of the chained outputs, so I looked for alternate ways to improve the generation quality. I ended up making SuperCOT at this time, by combining parts of datasets from Sahil Chaudhary’s Code Alpaca, CMU Mellon’s NeuLab CoNaLa, Google’s FLAN (QED and Aqua), and Peng et. al.’s Alpaca GPT-4, mostly sourced from Qingsi Yi et. al’s Alpaca CoT project dataset
The resulting model worked much better with chained prompting for me, but the quality was still a far cry from the demonstrations in Langchain’s documentation. I stopped working on the extension, instead I started playing with parts of the framework in isolation, such as vector databases and making my own lightweight chained prompting wrapper library. In the meantime, apparently some users found the model was very good at producing NSFW content. I had made it a point to filter the refusals and bias from the dataset as best as I could, so it was not unexpected. Soon after, the idea was floated to make models based on solely online roleplay logs, with the idea that such models would be much better at making chatbot outputs and would require no filtering of refusals. So, I started working on SuperHOT, and in the meantime others also worked on their own models, such as Bluemoon 13/30B.