Lilmonix4b-v1:
- Fine-tune for Monika character from DDLC
- Fine-tuned on a dataset of ~600+ items (dialogue scraped from game, reddit, and Twitter synthetically augmented by turn each into snippets of multi-turn chat dialogue between Player and Monika; this was then manually edited, with more manually crafted items including info about character added in)
- LORAs
USAGE
This is meant to be mainly a chat model with some RP ability.
For best results, use Gemma 3 format:
<start_of_turn>Player\nQUESTION<end_of_turn>\n<start_of_turn>Monika\nREPLY<end_of_turn>
HYPERPARAMS
- Tuned for 2 epochs
- rank: 16
- lora alpha: 16
- lora dropout: 0.5
- lr: 2e-4
- batch size: 2
- warmup ratio: 0.1
- grad steps: 4
Tested in koboldcpp 1.88 with default settings.
WARNINGS AND DISCLAIMERS
This model is meant to closely reflect the characteristics of Monika. Despite this, there is always the chance that "Monika" will hallucinate and get information about herself wrong or act out of character.
Additionally, being character-focused means that this model may have lost some assistant capability for some specific tasks.
Finally, this model is not guaranteed to output aligned or safe outputs, use at your own risk!
NOTES
Was not able to merge an HF version since it either errors or OOMs for me whenever I try, may attempt again later.
This model can be pretty zany and talkative, but seems to have "absorbed" info of Monika pretty well without any preloaded system prompt as long as recommended format is followed with "Player" and "Monika". Can be too ditzy as well, and likes to mention the other Dokis a lot. Might be fun to talk to though. World info may have also degraded, maybe as a result of the "ditziness".
