Training
The process of teaching an AI model by exposing it to data and adjusting its parameters to minimise errors.
In-Depth Explanation
Training is the process of teaching an AI model by showing it examples and adjusting its parameters to improve performance. For language models, this involves learning to predict text from massive datasets.
Training stages for LLMs:
- Pre-training: Learning general language patterns from vast text data
- Fine-tuning: Adapting to specific tasks or domains
- RLHF: Aligning behaviour with human preferences
- Continued training: Ongoing improvements and updates
What happens during training:
- Model processes training examples
- Compares predictions to actual targets
- Calculates error (loss)
- Backpropagates error to adjust weights
- Repeats billions of times
Training requirements for LLMs:
- Data: Trillions of tokens of text
- Compute: Thousands of GPUs for months
- Cost: $10M-$100M+ for frontier models
- Expertise: Large teams of ML engineers
This is why most businesses use pre-trained models rather than training from scratch. Fine-tuning offers a middle ground - smaller data and compute requirements to specialise an existing model.
Business Context
Training large models from scratch costs millions. Most businesses use pre-trained models, optionally with fine-tuning for specialisation.
How Clever Ops Uses This
We help Australian businesses leverage pre-trained models effectively. Full training is rarely needed - fine-tuning or RAG typically achieves business goals at a fraction of the cost.
Example Use Case
"GPT-4 was trained on trillions of tokens of internet text, books, and code, requiring months of compute on thousands of GPUs."