Gigax/NPC-LLM-3_8B: Domain-Specific NPC AI for Games

Gigax/NPC-LLM-3_8B is a specialized language model developed by Gigax, fine-tuned from Microsoft's Phi-3 architecture using LoRA. Its primary purpose is to power Non-Player Characters (NPCs) in games by generating contextually relevant actions and dialogue.

Key Capabilities

• Structured Command Generation: The model excels at parsing detailed game scene descriptions and outputting specific, actionable commands in a <action> <param> format (e.g., say <player> "Hello", greet <player>, attack <player>).
• Domain-Specific Training: It has been specifically trained to understand and respond within game contexts, leveraging inputs like world knowledge, known locations, NPC details, protagonist profiles, memories, quests, and allowed actions.
• Guided Generation: Designed to work seamlessly with libraries like outlines, it enables guided generation to ensure outputs strictly adhere to predefined formats, crucial for game logic integration.
• Overfit for Prompt Format: The model is intentionally overfit on a specific input prompt format, requiring users to follow this structure closely for optimal performance and predictable outputs.

Ideal Use Cases

This model is particularly well-suited for game developers looking to implement:

• Dynamic NPC Behavior: Creating NPCs that react intelligently and contextually to game events and player actions.
• Interactive Dialogue Systems: Generating natural and relevant dialogue for NPCs based on the game state.
• Automated Game Logic: Translating complex scene descriptions into executable game commands, streamlining development of interactive elements.

Gigax provides a Python client library to simplify integration and usage, handling prompt construction and output parsing.