AndriLawrence/Qwen-3B-Intent-Microplan-v1

Qwen-3B-Intent-Microplan-v1: Deprecated Game NPC Model

This model, developed by AndriLawrence, is a 3.1 billion parameter supervised fine-tune of bunnycore/Qwen2.5-3B-RP-Mix. It was the initial release of the "Intent-Microplan Framework," aiming to create dynamic, local-first, real-time game NPC brains by separating high-level social goals (intent) from low-level physical actions (microplan). The model outputs strict, engine-parsable JSON for dialogue and action.

Key Capabilities (Intended):

• Structured JSON Output: Designed to provide dialog, intent, and microplan in a strict JSON format for game engine integration.
• Dynamic Behavior Generation: Enables NPCs to generate plans based on context, facilitating emergent behavior.
• Local-First Deployment: Available in FP16 Transformers weights and quantized GGUF files for llama.cpp, suitable for in-game deployment (Unity, Unreal).
• Specific Intent Recognition: Trained to identify intents like social_greeting, acknowledge_touch, comfort_intimate, and react_to_player_action.

Why it's Deprecated:

This v1 model is considered a failure and is deprecated. The primary issue was the base model, bunnycore/Qwen2.5-3B-RP-Mix, which proved highly resistant to strict JSON schema enforcement, often breaking out of the required format. It has been superseded by Qwen-3B-Intent-Microplan-v2, which uses a more suitable foundational base model for structured data output.

Training Details:

• Base Model: bunnycore/Qwen2.5-3B-RP-Mix
• Method: PEFT LoRA fine-tuning on a custom English-only dataset focused on comfort and companion interactions.
• Context Length: 32768 tokens.