DQN-GPT-v0.2

DQN-GPT v0.2

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.

⸻

Motto

DQN Labs — Local AI for Everyone.