DQN-Labs/dqnCode-v0.2-1.5B

🧠 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