This is a fine tuned GPT-2 model trained on a corpus of all available English language Magic the Gathering card flavour texts.
This is intended only for use in generating new, novel, and sometimes surprising, MtG like flavour texts.
1from transformers import GPT2Tokenizer, GPT2LMHeadModel
2
3tokenizer = GPT2Tokenizer.from_pretrained("rjbownes/Magic-The-Generating")
4
5model = GPT2LMHeadModel.from_pretrained("rjbownes/Magic-The-Generating")
6
The training corpus was surprisingly small, only ~29000 cards, I had suspected there were more. This might mean there is a real limit to the number of entirely original strings this will generate.
This is also only based on the 117M parameter GPT2, it's a pretty obvious upgrade to retrain with medium, large or XL models. However, despite this, the outputs I tested were very convincing!
The data was 29222 MtG card flavour texts. The model was based on the "gpt2" pretrained transformer:
https://huggingface.co/gpt2.
Only English language MtG flavour texts were scraped from the
Scryfall API. Empty strings and any non-UTF-8 encoded tokens were removed leaving 29222 entries.
This was trained using google Colab with a T4 instance. 4 epochs, adamW optimizer with default parameters and a batch size of 32. Token embedding lengths were capped at 98 tokens as this was the longest string and an attention mask was added to the training model to ignore all padding tokens.
Average Training Loss: 0.44866578806635815.
Validation loss: 0.5606984243444775.
-
"Every branch a crossroads, every vine a swift steed."
—Gwendlyn Di Corci
-
"The secrets of this world will tell their masters where to strike if need be."
—Noyan Dar, Tazeem roilmage
-
"The secrets of nature are expensive. You'd be better off just to have more freedom."
-
"Even the Dwarves knew to leave some stones unturned."
-
"The wise always keep an ear open to the whispers of power."
1@article{BownesLM,
2 title={Fine Tuning GPT-2 for Magic the Gathering flavour text generation.},
3 author={Richard J. Bownes},
4 journal={Medium},
5 year={2020}
6}
7
