Llama 3.2 1B Fine-Tuning (Instruction SFT)
Overview
This project demonstrates fine-tuning of Meta Llama 3.2 1B using supervised fine-tuning (SFT) for instruction-following tasks. The objective was to adapt the base model to produce domain-relevant responses while maintaining the general reasoning ability of the original Llama architecture.
The training was performed using parameter-efficient fine-tuning (PEFT) with LoRA adapters to reduce compute cost and memory consumption while achieving strong performance.
Base Model
- Model: Meta Llama 3.2 1B
- Architecture: Transformer-based causal language model
- Training Method: Supervised Fine-Tuning (SFT)
- Fine-Tuning Technique: LoRA (Low-Rank Adaptation)
Hardware Setup
Training was executed on:
- GPU: NVIDIA RTX 4090 (24GB VRAM)
- CUDA Support: Enabled
- Precision: Mixed precision (FP16 / QLoRA supported)
- Training Time: ~110 minutes (approximate, varies by configuration)
The RTX 4090 provides excellent throughput for parameter-efficient tuning, enabling large model experimentation on consumer-grade hardware.
Dataset
The training dataset consists of instruction-response pairs structured as:
{
"instruction": "...",
"response": "..."
}
