Euterpe V2 has arrived, horns blaring!
The V2 finetune iteration of our Euterpe AI Model, including training for Custom AI Modules, is now available to all subscription tiers and the Free Trial.
Euterpe V2 is the result of the second iteration of the NovelAI finetune, aiding the model in improving coherence and overall comprehension.
Additionally, Opus tier users will continue to be able to select Euterpe V1 in the AI Model dropdown menu. This will let you compare the performance changes first hand.
We plan to phase out Euterpe V1 in the near future, so make sure to compare the models while you can!
On top of that, we've integrated Euterpe into the Free Trial and reset the free trial actions so everyone can test our strongest AI model - even if they've already used up their free trial actions testing with our previous model, Sigurd.
So, if you are not subscribed already, go and give Euterpe a spin - No account or credit card information needed!
The wait is over, Custom AI Module training for Euterpe has arrived!
If you've already created modules for Sigurd, with Custom AI Modules for Euterpe you will notice a more powerful effect with fewer steps.
Training modules for either AI Model cost the same amount of steps, but you might need less steps for Euterpe's AI Module training, since Euterpe responds much stronger to Custom AI Modules.
Give Euterpe that extra flair or knowledge by training a module with your favorite material!
With a new finetune come new logit probabilities, so we've included new and updated Default Setting Presets for Euterpe V2:
Genesis, Basic Coherence, Ouroboros, Ace of Spades, Moonlit Chronicler, All-Nighter, Low Rider and Morpho
Thanks to our contributors that aided in creating our new presets.
Keeping Euterpe V2's new logit probabilities in mind, your old Phrase Biases might need to be adjusted to get the same effects.
We recommend increasing the bias numbers since Euterpe requires much higher values for you to truly see the effect of your Biases.
We've also added the new Order Settings option, which can be found under Advanced Options. Each Setting option used to be performed in a specific sequence, influencing the tokens the sampling option considered.
Now you have complete control over the order in which your Story Settings are applied, giving you an array of different effects and possibilities.
We highly recommend advanced users to experiment with it!
You may have noticed the new Settings choices above, Top A and Typical Sampling!
Top A: Allows less variation when the AI is certain and allows more when it is uncertain.
Typical Sampling: Pick tokens by amount of information contained in them, allowing it to even discard more likely tokens in favor of less likely ones to better approximate human speech.
A final update regarding token bans and phrase biases.
In the interest of protecting NovelAI against unreasonable utilization - slowing down generations for all users - the following limits have been instated:
Bans are limited to no more than 2048 bans.
Biases are limited to no more than 1024 biases.
Should you exceed these amounts, you will receive an error message and will need to adjust the amount of biases or bans you have set.
However, Euterpe is not the only AI Model in this update that gets a little TLC!
Genji V2, our Japanese Storyteller model received a new, significantly improved tokenizer and a new finetune trained on a much more diverse body of Japanese text.
The new round of training also readily helps the model understand story tags and genre information.
Using Genji V2's new tokenizer update, you'll be able to utilize your context space much more efficiently!
Last but not least we wanted to take a second to show off our latest side project: GooseAI
Anlatan (that's us!) has teamed up with CoreWeave to provide a fully managed API service for developers who are creating the next generation of AI products.
Take advantage of $10.00 free credit to explore our Goose.ai playground and take flight with your own AI project. It's also a great way to take a gander at the newly released NeoX 20B AI model!