NovelAI

NovelAI is a subscription-based dynamic storyteller hosting fine-tuned GPT models. Run by a Turk.

Tiers
NovelAI has a free trial and offers three subscription tiers.

The free trial works like this:


 * 1) Before registration, you are allowed 50 actions to generate AI responses with.
 * 2) Once they run out, you can register an account with the service and verify your email to net a further 50 actions.
 * 3) You can't generate AI responses once you've run out of these actions. The next step would be to select a tier to subscribe to.

The three subscription tiers are as follows:


 * $25 Opus Tier - Unlimited Priority Actions, 2048 Token Context Window, 600 Characters Max Output Length, Experimental Features, 8000 Steps per month to train custom modules.
 * $15 Scroll Tier - 1000 Max Priority Actions each week, 2048 Token Context Window, 100 Max Output Token Length, 500 steps per month to train custom modules.
 * $10 Tablet Tier - 1000 Max Priority Actions each week, 1024 Token Context Window, 100 Max Output Token Length, 500 steps per month to train custom modules.

Scroll tier is the best to go for if you're just starting the service. The extra 1024 tokens make a huge difference in play compared to just half.

NovelAI also comes with the ability to purchase perpetual module training steps:


 * $3.79 for 2,000 steps
 * $6.49 for 5,000 steps
 * $10 for 10,000 steps

As said before, these purchased steps persist until you use them.

Models
Currently NovelAI has multiple: Calliope, Sigurd, and Euterpe. All are fine-tuned on their own custom dataset.

Calliope is based on GPT-Neo (2.7B parameters), Sigurd is based on GPT-6B-J, and Euterpe is based on Fairseq 13B. Sigurd is the name of NovelAI fine-tuned GPT-J-6B model. This model launched during the Beta as an Experimental Feature for our opus tier subscribers and has since been rolled out to all tiers. Newer Sigurd models will continue to release as Experimental Feature for Opus subscribers first. Sigurd is a much bigger model and response times are a little slower in comparison to Calliope, our 2.7B model.

Euterpe is based on the Fairseq 13b model—hosted and running on the NovelAI infrastructure, like all other models. Euterpe is the latest Experimental AI model to join the NovelAI AI model family. Euterpe does not support custom AI modules yet but had the entire catalog of all Default AI modules available on release. Euterpe was released as V0 with a 65% epoch finetune on January 10th, 2022. Euterpe V1 was added on January 11th, 2022, with a completed first epoch and 7% finetune on the second epoch.

Finetuning
While the specifics have yet to be revealed by any member of the team, but we do have a general understanding. Kurumuz and his colleagues stated in the early days of the project that they planned to work primarily with books, light novels, and high-quality novellas for their dataset.

The first version of the set was used to fine-tune Calliope, then Sigurd (5% at V1, then completed at V2), but there was an oversight regarding dialogue quality, which necessitated a dataset pruning to remove offending stories. After that, they fine-tuned again, resulting in Sigurd V3. Finally, they expanded and improved the dataset to culminate in the 4th iteration of Sigurd. It's believed that the dataset is now >10 GB in size, though new text is constantly added to the dataset by the fine-tuning team.