Model Overview
AndriLawrence/Qwen-3B-Intent-Microplan-v2 is a 3.1 billion parameter model, fine-tuned from Qwen2.5-3B-Instruct, specifically for intent and microplan-driven NPC dialog in VR and game environments. It processes a structured CONTEXT JSON (including environment, relationship, mood, and signals) and outputs a strict JSON object containing dialog, intent (selected from 19 predefined labels), and microplan (action primitives).
Key Capabilities
- Strict JSON Output: Guarantees a single JSON object with
dialog, intent, and microplan, crucial for programmatic control. - Contextual Understanding: Interprets detailed game state via a
CONTEXT JSON to inform responses. - NPC Logic & Dialog Generation: Provides both the underlying intent and the surface-level dialog for NPCs.
- Persona Adherence: Designed for consistent character tone and behavior, with an example 'Rin' persona provided.
- Optimized for Real-time: Suitable for low-latency applications like game companions.
What Makes This Model Different?
Unlike general-purpose LLMs, this model is hyper-specialized for a very specific task: generating structured, actionable responses for game NPCs. Its primary differentiator is the guaranteed strict JSON output for intent and microplan, which allows developers to integrate LLM-driven dialog directly into game logic. It's not just about generating text; it's about generating controllable game actions and character states alongside dialog. The v2 iteration features improved JSON guardrails, better label alignment, and enhanced persona consistency, making it more stable and reliable for production use.
Recommended Use Cases
- VR/Game NPC Systems: Ideal for creating dynamic and responsive non-player characters.
- Two-Stage Dialog Pipelines: Can serve as the first stage (intent/microplan) before a persona-specific dialog model.
- Custom Character Development: The base SFT checkpoint is an excellent starting point for fine-tuning new character personas while retaining the core JSON output structure.