LLM Fine-Tuning Lab

Overview

The LLM Fine-Tuning Lab is the crown jewel of our lab infrastructure. Each student gets a dedicated NVIDIA A100 with 80 GB of VRAM — the highest-memory GPU available on GCP — giving you enough room to fine-tune 7B parameter models with full precision or run QLoRA on models up to 70B parameters. This is the same hardware that AI companies use to create their custom language models. The 1 TB SSD ensures fast model loading — downloading a 70B model from Hugging Face takes minutes, not hours. Pre-installed tooling includes Unsloth for 2x faster fine-tuning, Axolotl for config-driven training, and Flash Attention 2 for memory-efficient attention computation.

What You'll Do in This Lab

• Fine-tune Llama 3, Mistral, and Gemma models on domain-specific datasets
• Apply LoRA and QLoRA adapters for parameter-efficient fine-tuning
• Fine-tune 70B parameter models in 4-bit quantization on a single GPU
• Create instruction-tuned chat models from base models
• Implement DPO alignment to steer model behavior with preference data
• Merge LoRA adapters back into base models for deployment

Lab Workflow