Thespis-34b-DPO-v0.7

Thespis v0.7 ( Yi 34b Finetune - 8k ctx)

This model works best with internet style RP using standard markup with asterisks surrounding actions and no quotes around dialogue. Current work is looking at creating a custom DPO dataset as well as expanding the training data to include more niche interests.

External Datasets Used:

• Pure-Dove Dataset
• Claude Multiround 30k
• OpenOrcaSlim
• Augmental Dataset

DPO was done using a few generic datasets available on Hugging Face.

DPO Data:

• Intel/orca_dpo_pairs
• NobodyExistsOnTheInternet/ToxicDPOqa

Works with standard chat format for Ooba or SillyTavern.

Prompt Format: Chat ( The default Ooba template and Silly Tavern Template )

{System Prompt}

Username: {Input}
BotName: {Response}
Username: {Input}
BotName: {Response}

Ooba ( Set it to Chat, select a character and go. )

Silly Tavern Settings ( Default )

## License

As a model based on Yi, this finetune is under the Yi license.