Attention (new)
Version 2 is out! https://huggingface.co/TheDrummer/Moistral-11B-v2
Attention (old)
I have created a better version here: https://huggingface.co/TheDrummer/Moistral-11B-v1.1-MERGED-PCT-GGUF
Moistral 11B v1 💦
An eRP model with a rich vocabulary of moist. Finetuned by yours truly.
GGUF: https://huggingface.co/TheDrummer/Moistral-11B-v1-GGUF/tree/main
Description
Moistral 11B is an aggressive finetune of the legendary Fimbulvert v2 model.
If you haven't heard of that model, then your poor-GPU ass is missing out! It is an amazing substitute for those heavyweight models, and it won't disappoint. (That's coming from someone who daily drives a Venus / Goliath 120B)
...and if you're looking for something more moist, then you've come to the right place.
You can expect the same sort of performance here! It's the Honda Civic of eRP models!
Pimp my ride
Turn any old story into a Moistral masterpiece. In fact, we recommend that you leave out any hints of moist in your story to let Moistral cook.
Drive it like it's hot
When in Instruct Mode, Moistral will treat you like a director of your own fantasy ride. Be a controlling navigator or a passenger princess, it won't matter - you're in for a good ride!
It's got the mileage
Moistral has been trained with many long form texts, a nice chunk of which are 8K in length. It is capable of going far and long without passing it back to you. This is not your typical chibi RP model.
Parameters
If Moistral starts to underperform and token spit, I've noticed that lowering the parameters makes it coherent again. Here's what worked for me:
temperature: 0.66
repetition_penalty: 1.1
top_p: 0.64
rp_slp: 1
I encourage you to play around with the parameters yourself to see what works for you.
What's next?
Moistral 11B is my first attempt at finetuning a capable model (Sorry, CreamPhi-2). It's coherent and creative enough to let me understand the impact of my dataset & training. Playing around with it has already given me a better idea on the Do's and Don'ts. I will most likely make a version 2 with some improvements:
- Remove any glitchy texts that come from my dataset. Sanitize sanitize sanitize!
- Balance out the themes in the dataset for a richer, more diverse experience.
- Consider extending the context window.
- Add a 'monologue' dataset that forces the model to keep talking without much interaction from the
user. - Maybe, just maybe, expose it to dry stuff to let Moistral cook.
GGUF: https://huggingface.co/TheDrummer/Moistral-11B-v1-GGUF/tree/main
I have to acknowledge that I'm standing on the shoulders of giants. Thank you Sao for sharing your finetune config along with tips on getting started. Thanks to everyone in the Finetuning channel for entertaining my every question.
