This is incredible! Many potential utilizations if tuned.

#1
by bobbyhiddn - opened

This is the tool I've been looking for when it comes to generating unique spells for use in a procedural roguelike! I am still new to these concepts, but I want to take this tool and learn how to incorporate it into item-generation and spell-crafting systems.

Thanks to the creator of it for the inspiration and affirmation that it is possible!

Any advice on how I could go about tuning it for those purposes?

Hey! I'm so glad you like it. I fine-tuned a gpt-2 model on the spell names and descriptions of D&D 5e spells. I fine-tuned the model in December 2021 but didn't upload it until recently.
The format I used for training was

<|name|> SPELL_NAME <|spell|> SPELL_DESCRIPTION

When I prompt the model I use

<|name|> SPELL_NAME <|spell|>

and have it generate the rest.

Like you, I eventually wanted to use the same model to generate items as well. I would have done something like
<|name|> ITEM_NAME <|item|> ITEM_DESCRIPTION
I don't remember what dataset I downloaded, but I just googled for .json files of D&D 5e spells.
I can't seem to follow the fine-tuning example I used, but I did find this one that seems relevant: https://colab.research.google.com/drive/13dZVYEOMhXhkXWfvSMVM1TTtUDrT6Aeh

If you have any questions, I'd be happy to try and help! :) And when you get one fine-tuned I'd love to see it!

You did a great job, and honestly, I’m trying to figure out step one.

Did you tune it in Tensorflow or PyTorch or am I completely off base?

Yeah, I used Pytorch and the transformers library. The Colab notebook I linked before isn't the exact process I used last time, but it's incredibly close and would be an excellent place to get started once you have the dataset you want to use. Next time I'm fine-tuning a model I will likely use that Colab notebook as a template to start from and modify it from there.

Sign up or log in to comment