Home

Welcome to TR's NAI Research Wiki!

Here you will find mainly two things:

• A record of my own findings and thoughts on research methodology from tinkering with Novel AI
• An archive of community research findings from the #community-reserach channel in the NAI discord for the period of 14/06/2021 to 04/07/2021.

Log of TR's research findings

I will post methodology and findings of my own experiments with NAI's models here - see sidebar to the right. I am trying to focus either on methods that are easy to replicate or on collecting data that could still be of use for later models.

I feel like this is the most reasonable path forward. Some things will have to be researched from scratch for new models (like A/N keywords) and others will have to be adjusted slightly (like finding the best amount of steps for module training). For the former, easily replicable methods should help. For the latter, collecting data for different models might help us to see patterns so that we get a feel for how certain things have to be adjusted for different models.

Archive of NAI community research findings (14/06/2021 to 04/07/2021)

During the early days of NovelAI, me and others tried to pull together the key findings from people's research efforts in #community-research in the NAI discord.

For the most part this was done by working through the chat log, skimming the messages from the date of open beta (14/06), trying to summarize what look like the key findings. These findings concern mostly Sigurd v2 and Sigurd v3, so it is unclear how valid those are for newer NAI models. Findings and observations for older NAI models do not necessarily carry over to newer ones. Also, these observations are not to be taken as gospel - confirming other people's observations from using NAI is hard and time-consuming! So treat most things in the archives as very tentative findings! In the few cases, where more extensive testing was done, and anecdotal observations held up, it is noted.

I think the most useful use of this archive should be to serve as inspiration for your own experimentation with newer models. Some stuff in there might work for all models, some of it might not, some of it might have been in mistake and was never true for any model. Either way, hopefully by trying things out with the newer models, we can all learn a bit more.