🧠 Qwen2.5 3B Instruct – Fine-Tuning on Custom Data

Fine-tune the Qwen2.5 3B Instruct model (by Alibaba) on your own instruction–response dataset using LoRA (Low-Rank Adaptation) for efficient, low-resource training.

This notebook provides a full, ready-to-run pipeline for domain adaptation, custom chatbot creation, or style-specific instruction tuning.

---

🚀 Features

✅ Fine-tunes Qwen2.5-3B-Instruct (3B parameters) using LoRA adapters
✅ 4-bit quantization (BitsAndBytes) for efficient memory use
✅ Works with custom instruction–response datasets (JSON, CSV, or text)
✅ Easy to train on Google Colab or local GPU
✅ Exports a ready-to-use fine-tuned model compatible with transformers pipeline

---

📖 Dataset Format

Your dataset should contain instruction–response pairs, for example:

[
  {
    "instruction": "Explain quantum computing simply",
    "output": "Quantum computing uses qubits that can represent 0 and 1 at the same time..."
  },
  {
    "instruction": "Write a short poem about AI",
    "output": "Machines that dream and think so bright..."
  }
]

⚙️ Training Configuration

Parameter	Description	Default
model_name	Base model	Qwen/Qwen2.5-3B-Instruct
r	LoRA rank	16
lora_alpha	Scaling factor	32
target_modules	Layers adapted	["q_proj", "v_proj"]
lora_dropout	Dropout	0.05
learning_rate	Learning rate	2e-4
epochs	Training epochs	3
batch_size	Per device batch size	1
max_length	Token limit	1024

🧱 References

• Qwen2.5 Models – Alibaba Cloud
• PEFT: Parameter Efficient Fine-Tuning
• Transformers Documentation
• BitsAndBytes
• Unsloth