Training Type of LLM

[Aidenzich]

Category	Pre-training	Post-Pretraining	Fine-tuning	Instruct-tuning
Definition	The initial training phase where the model learns from a large and diverse dataset	Further training after pre-training on a specific sub-domain or specialized dataset	Further training on a pre-trained and post-pretrained model using a smaller, task-specific dataset	Training the model to better follow instructions in prompts, focusing on improving instruction parsing and response
Purpose	To develop a broad understanding of language, context, and various types of knowledge	To refine the model's knowledge in specific sub-domains	To enhance the model's performance in specific scenarios or tasks	To improve the model's ability to parse and execute given instructions to align with user intentions
Dataset Size	Very large, often containing trillions of tokens	Large but focused on specific sub-domains	Smaller, focused on specific tasks or domains	Depends on the need to train the model to understand and execute instructions, dataset size varies
Computational Resources	Extremely high, often requiring millions of dollars	High, but lower than pre-training	Relatively low, as the dataset is smaller and more focused	Depends on the model and dataset, generally less than pre-training
Knowledge Expansion	Broadly expands the model's general knowledge	Expands and refines the model's knowledge in specific sub-domains	Enhances the model's performance in specific contexts	Does not add new factual knowledge, but improves the model's ability to parse and respond to prompts
Examples	Large language models learning from diverse web text	Further training on specialized literature in domains such as medicine, law	Fine-tuning GPT-2 to generate lyrics in the style of a specific artist, such as Eminem	Training models like ChatGPT and InstructGPT to better understand and execute user instructions
Users	Large companies and research institutions; beginners will not be involved in pre-training	Researchers and developers in specialized fields, further refining model knowledge	Can be done in various research and application contexts	Used for applications and services where the model needs to understand and execute specific instructions, such as ChatGPT