🧠 DQN Code v0.2
A 1.5B Python Specialist
DQN Code v0.2 is a lightweight coding-focused model built on Qwen2.5-1.5B-Instruct and fine-tuned specifically for high-quality Python generation.
This release focuses on algorithmic correctness, structured implementation, and clean function completion.
🚀 Highlights
- 🔹 1.5B parameters
- 🔹 LoRA fine-tuned
- 🔹 Python-specialized
- 🔹 Optimized for deterministic completion
- 🔹 Designed to run locally on 8GB systems
📊 Benchmark Performance
HumanEval (0-shot, temperature=0.0)
Evaluated using mlx_lm.
| Model | Parameters | HumanEval pass@1 |
|---|
| Qwen2.5-1.5B-Instruct (official) | 1.5B | ~37.8% |
| MistralAI Mistral 7B | 7B | ~30.5% |
| Google Gemma 2 9B (as on llmstats.com) | 9B | ~40.2% |
| DQN Code v0.2 | 1.5B | 49.39% |
Evaluation settings:
- 0-shot
- Temperature = 0.0
- No few-shot prompting
- Full 164 HumanEval tasks
This represents a +11.6% absolute improvement over the base Instruct model.
🎯 Design Philosophy
Instead of scaling parameters, DQN Code focuses on:
- High-quality distilled supervision
- Python-heavy training distribution
- Clean function-style completions
- Reduced conversational overhead
- Local inference efficiency
Small models benefit heavily from specialization.
This release demonstrates how targeted fine-tuning can significantly improve coding performance without increasing model size.
🔧 Training Details
- Base:
Qwen2.5-1.5B-Instruct
- Fine-tune type: LoRA
- Effective batch size: 8
- Max sequence length: 512
- Optimizer: AdamW
- Learning rate: 5e-6
- Dataset size: ~1.8k curated Python-focused samples
- Training hardware: 8GB RAM system (MLX)
Training focused on:
- Function completion
- Algorithmic correctness
- Clean Python structure
- Reduced hallucinated commentary
💻 Intended Use
- Local coding assistant (with decent performance in other languages too!)
- Python function completion
- Algorithm practice
- Educational use
- Lightweight code generation
⚠ Limitations
- Primarily optimized for Python (but performs well on other languages too.)
- Not benchmarked on multi-language coding
- Limited evaluation on mathematical reasoning
- Not trained for tool use or multi-step planning
🌍 Philosophy
Powerful coding models do not require massive infrastructure. You don't need a datacenter at home!
Focused training + efficient inference can deliver strong results on modest hardware.
❓ Queries
If you have any questions regarding the model, want to know how it was trained and our pipleine process, how you can make a better version of the model, or you just want to chat about AI, feel free to contact me on Discord at @dqnlabs.
Enjoy this model, for this is the best so far. There's more coming.
-- DQN Labs
This model card was made with the assistance of dqnGPT-v0.1-7B :D
