A LoRA fine-tune of SmolLM3-3B for D&D NPC roleplay, trained on character dialogue
drawing from Forgotten Realms, Lord of the Rings, and The Witcher lore, settings,
and characters.
Intended Use
Interactive NPC roleplay in tabletop RPG contexts. The model responds in-character
to player actions as directed by a Dungeon Master.
Required Prompt Format
The model requires a specific input format to trigger roleplay behavior:
System prompt:
You are an NPC actor in an ongoing D&D campaign. You embody characters as directed
by the Dungeon Master. Always structure your response using this format:
Physical action or reaction in italics.
"Spoken dialogue in quotes."
Expected output format:
Grabs the edge of the bar and leans forward slowly.
"You'd best be careful who you ask that question, stranger."
Training Details
Base model: HuggingFaceTB/SmolLM3-3B
Fine-tune method: LoRA / QLoRA
Dataset: Synthetically generated D&D NPC dialogue with lore grounding from
Forgotten Realms, LOTR, and Witcher universes, mixed with PygmalionAI/PIPPA
and Norquinal/claude_multiround_chat_30k
