VLM Series

0.1 Series Overview

This series takes a fastai-inspired, top-down approach to Vision-Language Models (VLMs). We’ll start with real-world applications and working code, then dive deeper into the concepts behind these powerful models. The focus is on open-source VLMs and practical workflows using PyTorch and HuggingFace.

0.1.1 Learning Philosophy

• Start with working code: Each post begins with hands-on examples defining real tasks
• Iterative deepening: Concepts are revisited with increasing depth
• Learn by building: Notebooks and exercises included for every topic

0.1.2 Posts in this Series

1. Hands-On with Qwen3-14B: A Reasoning-Centric LLM for Data Scientists
   • What is Qwen3-14B?
   • How to use it for vision-language tasks
   • Hands-on: inference and fine-tuning
2. Fine-Tuning Qwen3-14B with Unsloth
   • Practical guide to fine-tuning with Unsloth
   • Colab and local multi-GPU workflows
   • Tips for data scientists

0.2 Prerequisites

• Basic Python and PyTorch
• Familiarity with Jupyter notebooks
• Some experience with deep learning (helpful, not required)

0.3 Tools We’ll Use

• PyTorch for model development
• HuggingFace Transformers for VLMs
• FastAI for rapid prototyping (where applicable)
• Open-source VLMs (Qwen, LLaVA, etc.)

0.4 Getting Help

• Use the comments section below each post
• Check the GitHub repository for code
• Join our discussion forum (coming soon)

Note

This series is updated regularly based on reader feedback and new developments in VLM research.