Introduction to LLM Fine-Tuning
Chapter 1: Foundations of Model Customization
What is Fine-Tuning?
Pre-training vs. Fine-Tuning
When to Fine-Tune: An Analytical Framework
Overview of Fine-Tuning Strategies
The Role of Transfer Learning in LLMs
Setting Up Your Development Environment
Chapter 2: Data Preparation for Fine-Tuning
Sourcing and Selecting High-Quality Datasets
Instruction-Based vs. Conversational Data Formats
Data Cleaning and Preprocessing Techniques
Creating and Structuring Custom Datasets
Tokenization for Fine-Tuning
Hands-on Practical: Building a Fine-Tuning Dataset
Chapter 3: Full Parameter Fine-Tuning
The Mechanics of Full Fine-Tuning
Architectural Considerations for Full Fine-Tuning
Managing Computational Resources
Configuring Training Arguments and Hyperparameters
Monitoring Training: Loss and Metrics
Saving and Loading Fine-Tuned Models
Practice: Full Fine-Tuning on a Small-Scale Model
Chapter 4: Parameter-Efficient Fine-Tuning (PEFT)
Introduction to Parameter-Efficient Fine-Tuning
Low-Rank Adaptation (LoRA): Theory and Operation
Implementing LoRA with the PEFT Library
Quantization and its effect on Fine-Tuning (QLoRA)
Other PEFT Methods: A Brief Survey
Comparing PEFT and Full Fine-Tuning Trade-offs
Hands-on Practical: Fine-Tuning with LoRA
Chapter 5: Evaluation and Deployment
Defining Performance Metrics for Generative Tasks
Quantitative Evaluation: ROUGE, BLEU, and Perplexity
Qualitative Evaluation: Human-in-the-Loop Assessment
Building an Evaluation Pipeline
Strategies for Merging Adapters with the Base Model
Preparing Models for Inference
Practice: Evaluating a Fine-Tuned Model
Managing Computational Resources
Was this section helpful?
- Mixed-Precision Training, Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, Hao Wu, 2018 International Conference on Learning Representations (ICLR) DOI: 10.48550/arXiv.1710.03740 - Introduces the concept of mixed-precision training (FP16), demonstrating its benefits for reducing memory and increasing speed while maintaining accuracy in deep learning models.
- Training Deep Nets with Sublinear Memory Cost, Tianqi Chen, Bing Xu, Chiyuan Zhang, Carlos Guestrin, 2016 arXiv DOI: 10.48550/arXiv.1604.06174 - Proposes gradient checkpointing (or activation checkpointing) as a technique to trade computation for memory, allowing training of deeper networks than previously possible.
transformers.TrainingArguments, Hugging Face, 2024 (Hugging Face) - Official documentation forTrainingArgumentsin the Hugging Face Transformers library, detailing parameters relevant to memory management likefp16,bf16,gradient_accumulation_steps,gradient_checkpointing, andoptim.