​
Levels of Training Large Language Models (LLMs)

In the world of LLMs, there are four primary levels of training and fine-tuning. Each level serves a specific purpose and involves different amounts of data and computational resources. Let’s explore these levels in detail.

​
1. Pretraining on Raw Text

Pretraining is the foundation of creating a new base model. This level is used when you want to develop a model from scratch or create a specialized base model for specific domains like mathematics, medicine, or other specialized topics.

​
Characteristics:

  • Data Volume: Involves training on enormous amounts of text data, often trillions of tokens.
  • Computational Resources: Extremely resource-intensive and expensive.
  • Duration: Can take weeks or months, even with powerful hardware.
  • Purpose: To learn general language patterns, grammar, and broad knowledge.

​
Example Use Cases:

  • Creating a new general-purpose LLM like GPT-3 or BERT.
  • Developing a domain-specific base model (e.g., BioBERT for biomedical text).

This is very rare, but if you want train a base model . You need to create a .jsonl file with the text chunk

Dataset Type: TEXT

dataset.jsonl
{"text": "Some important task with context" }
{"text": "Some important task with context" }
{"text": "Some important task with context" }

​
2. Instruction Tuning

Instruction tuning is the process of fine-tuning a pretrained model to follow specific instructions or prompts. This is the most common fine tuning, since base model has compressed most the world knowledge, you only have to train them on your task.

​
Example Use Cases:

  • Text-SQL, TEXT-API, Question Answering, Text to action, Retrievals, Generate HTML design from prompt
  • Enhancing the model’s understanding of task-specific prompts.
  • Fine tune small model on synthetic dataset from a bigger model for cost reduction.
  • Fine tune small model on to prevent long repetitive prompt and speed.
  • Fine tune big model on hard task that father modal cant solve.
dataset.jsonl
{"instruction": "Some important task with context ?", "output": "Boo" }
{"instruction": "Some important task with context ?", "output": "Boo" }
{"instruction": "Some important task with context ?", "output": "Boo" }

​
3. Chat Tuning

Chat tuning is a specialized form of instruction tuning that focuses on improving the model’s conversational abilities. So if build an interactive chatbot this is what you need to train

​
Example Use Cases:

  • Developing chatbots or virtual assistants.
  • Improving the model’s ability to handle multi-turn conversations and maintain context.
dataset.jsonl
[{"role": "user", "content": "Hi can I train llama model"}, {"role": "assistant", "content": "Sure"}]
[{"role": "user", "content": "Hi can I train llama model"}, {"role": "assistant", "content": "Sure"}]
[{"role": "user", "content": "Hi can I train llama model"}, {"role": "assistant", "content": "Sure"}]

​
4. Direct Preference Optimization (DPO)

DPO is an advanced fine-tuning technique that aims to align the model’s outputs with human preferences.

​
Example Use Cases:

  • Improving the model’s ability to generate more natural, ethical, or stylistically appropriate responses.
  • Fine-tuning the model to adhere to specific guidelines or tone preferences.
dataset.jsonl
{ "instruction": "....", "chosen": "...", "rejected": "..." }
{ "instruction": "....", "chosen": "...", "rejected": "..." }
{ "instruction": "....", "chosen": "...", "rejected": "..." }

​
Conclusion

Each level of LLM training serves a specific purpose in the development and refinement of these powerful AI models. While pretraining creates the foundation, the subsequent levels of fine-tuning allow for more specialized and targeted improvements. Understanding these levels can help in choosing the right approach for developing or customizing LLMs for specific applications.

QuickstartModels