DQN-GPT v0.2 is a 3.4B parameter instruction-following language model designed with a single guiding philosophy:
AI with obedience in mind.
The model prioritizes precise instruction adherence, reliable formatting, and predictable behavior for developers building local AI systems.
⸻
Model Overview
• Model Name: DQN-GPT v0.2
• Parameters: 3.4B
• Base Architecture: Extracted text backbone from Ministral 3 3B
• Training Type: Instruction fine-tuning
• Developer: DQN Labs
• License: Apache 2.0
DQN-GPT v0.2 was created by extracting the pure text model component from the Ministral 3 3B architecture. This custom extraction removed non-essential components and retained only the core language modeling capability.
The resulting base model was then instruction tuned on a carefully curated dataset of ~20,000 samples, focusing specifically on improving:
• Instruction obedience
• Response clarity
• Task completion reliability
• Structured outputs
This design philosophy emphasizes predictable and controllable AI behavior rather than pure conversational creativity.
⸻
Training Details
Base Model Extraction
The base model used for DQN-GPT v0.2 originates from Ministral 3 3B.
However, instead of using the full architecture, the text-only language model component was extracted and repurposed as a standalone foundation model. This approach allowed tighter control over the training pipeline and model behavior.
Instruction Fine-Tuning
The model was fine-tuned on a curated mix of approximately 20,000 instruction examples, combining several instruct-style datasets designed to improve:
• task completion accuracy
• formatting compliance
• structured responses
• deterministic instruction following
The dataset mixture was selected and filtered specifically to reinforce obedience and clarity rather than casual chat behavior.
⸻
Design Philosophy
Most instruction models aim for helpfulness and personality.
DQN-GPT v0.2 instead focuses on something different:
Obedience.
The goal is to create a model that:
• does exactly what the user asks
• respects formatting instructions
• follows structured prompts reliably
• avoids unnecessary verbosity
This makes the model particularly useful for:
• AI agents
• automation workflows
• coding assistants
• structured generation tasks
• tool-using systems
⸻
Intended Use
DQN-GPT v0.2 is designed primarily for:
• Local AI assistants
• Developer tools
• Coding help
• Instruction-following tasks
• Structured text generation
• Agent pipelines
The model performs best when prompts are clear and structured.
⸻
Limitations
• The model is relatively small compared to frontier models.
• Knowledge is limited by the base model’s training data.
• Performance on complex reasoning tasks may vary.
• Not supported for multimodal tasks.
