Qwen2.5 (3B & 7B) Reasoning Model Fine-Tuning Suite
This repository contains the complete technical architecture, dataset processing scripts, evaluation metrics, LoRA adapter weights, and executive PowerPoint presentation for Reasoning & Chain-of-Thought (<think>) Fine-Tuning of Qwen2.5-3B-Instruct and Qwen2.5-7B-Instruct across 2× NVIDIA H200 NVL GPUs.
Executive Scorecard & Method Benchmarks
1. Qwen2.5-3B Reasoning Benchmarks
| Fine-Tuning Method | Training Time | Peak VRAM | Training Loss | CoT (<think>) Adoption Rate |
|---|---|---|---|---|
| QLoRA (4-bit NF4) | 10.55 min | 14.50 GB | 0.942 | 100.0% (50/50) |
| LoRA BF16 (16-bit) | 5.58 min | 16.82 GB | 0.885 | 100.0% (50/50) |
| Full Fine-Tuning | 6.50 min | 42.10 GB | 0.812 | 100.0% (50/50) |
2. Qwen2.5-7B Reasoning Benchmarks
| Fine-Tuning Method | Training Time | Peak VRAM | Training Loss | CoT (<think>) Adoption Rate |
|---|---|---|---|---|
| QLoRA (4-bit NF4) | 19.77 min | 19.82 GB | 0.910 | 92.0% (46/50) |
| LoRA BF16 (16-bit) | 10.51 min | 22.45 GB | 0.841 | 96.0% (48/50) |
| Full Fine-Tuning | 12.11 min | 64.30 GB | 0.789 | 100.0% (50/50) |
Key Hardware & Velocity Insight: The QLoRA Speed Paradox
On enterprise hardware (2× NVIDIA H200 NVL with 288GB VRAM), 16-bit LoRA (10.51 min) and Full Fine-Tuning (12.11 min) trained ~1.9× faster than 4-bit QLoRA (19.77 min).
- Why: On high-bandwidth Tensor Cores, 4-bit NF4 weight dequantization creates CUDA kernel memory bottlenecks during backward passes. When VRAM is abundant, unquantized 16-bit LoRA provides maximum throughput.
Repository Structure
adapters/: PEFT LoRA adapter weights (adapter_model.safetensors&adapter_config.json) for 3B and 7B models.QWEN-3B_FINE_TUNING_FOR_REASONING/: Complete 3B code, docs, and metrics.QWEN-7B_FINE_TUNING_FOR_REASONING/: Complete 7B code, docs, and metrics.FineTuning_Qwen_3B_and_7B_for_Reasoning_Models.pptx: 12-slide widescreen executive presentation deck.