Amazing Model
Thanks for releasing this amazing model! It is the ONLY Nemo-based finetune that beats it in writing (for me personally). Its prose is much more human-like and virtually free of GPT-slop, even occasionally bypassing AI detection tools. I believe this is because of the dataset used?
However, there are a few flaws that are hindering the experience. Firstly, it can rush the story ending and finish it abruptly with 'THE END'. It struggles with continuing a story from where it left off, tending to completely rewrite it.
It would be interesting to see a gutenberg finetune of Rocinante v1 by TheDrummer which doesn't have these issues. It is the only other writing model I use.
Sure, I can give that a shot. Thanks for your feedback.
I'll agree as well. Though not perfect, your model definitely has a bit more edge to it in the way it conveys itself, and it's been really wonderful to use by itself and in my merge too. It keeps to the character well, delivers good prose, and knows how to expand inside the story. If you don't mind taking requests, then I'd like to throw my hat into the ring and ask if you have any interest doing the same treatment with Lyra-v1? It's the only other Mistral format model besides Magnum and Instruct that held up in some private testing for me and some others as well. ChatML models with Nemo have been shaky, to say the least, and that model is kinda a half-and-half with it's prompt template, using both it and Mistral. So I really think adding your training on top would help iron out it's quirks, even if just to make it more receptive to it's more stable Mistral prompting. But regardless, you've done good work, and I hope you labor continues to bear tasty fruits!
I'm gonna add that this model is extremely based and I wish more people didn't sleep on it. It writes very well and it has enchanced my merges a ton, not to mention, this one also works perfectly fine on higher contexts (at least, v1 does). Thank you for your work!
Rocinante base is done: https://huggingface.co/nbeerbower/mistral-nemo-gutenberg-12B-v4
Lyra-Gutenberg is in the oven.
Thanks everyone!