Show HN: Distil-NPC: a family of models for non-playable characters in games (github.com)

Distil-NPC is a small-family of specialised small language models fine-tuned to role-play non-playable characters (NPCs) in games by embedding character biographies directly into the model via a closed-book QA setup. The team trained models based on Google’s Gemma 270M and 1B variants and released them on Hugging Face, enabling natural-language NPC interactions without runtime retrieval or network calls. That on-device capability (the 270M model’s total size is ≈0.5 GB) addresses gaming constraints around latency and offline play, making real-time, free-text conversations with many NPCs practical for modern hardware. Technically, they preprocessed the amaydle/npc-dialogue dataset (81 character bios plus question–answer pairs) and trained a single model to answer questions for any character by prefixing inputs with the character name—avoiding one-model-per-character scaling. LLM-as-a-judge metrics show improved scores with scale (270M roughly 0.43–0.48; 1B roughly 0.58–0.61), and qualitative analysis finds finetuned models give more plausible, persona-consistent replies where base models sometimes echoed the prompt or produced random text. Implications: a compact, efficient route to richer, persistent NPC behavior and storytelling in games, though metrics can understate quality and further work is needed on consistency, safety, and broader evaluation.