Code and stuff for open-source AI Dungeon clone.
Check models and datasets here: https://huggingface.co/OpenDungeon
To train models use our version of CRD3 with train_opendungeon.py. You will need around 12GB of vRAM or around 8GB of vRAM if you will use 8bit mode.
Side-by-Side of base GPT-J with finetuned on CRD3:
| Like | Fluency | No Contradiction | Reaction | |
|---|---|---|---|---|
| basemodel | 46.02% | 44.57% | 48.10% | 44.90% |
| LoRA | 53.98% | 55.43% | 52.90% | 55.10% |
Answers are weighted according to probabilities of being correctly labeled provided by Toloka.
Demo: 
In alphabetic order:
Please, ignore the authorship of commits.