Training and Decoding of Large Language Models
Training and decoding are fundamental processes in the lifecycle of Large Language Models (LLMs). This chapter explores these processes in detail.
Training
Training an LLM involves exposing it to extensive text data and adjusting its parameters to improve performance in predicting and generating text.
Domain-Adaption
Domain-adaption refers to modifying a model to perform better on tasks or topics outside its original training domain. This often involves additional training on domain-specific data.
Training Styles
| Training Style | Modifies | Data | Summary |
|---|---|---|---|
| Fine-Tuning (FT) | All parameters | Labeled, task-specific | Traditional method involving complete re-training. |
| Parameter Efficient FT | Few, new parameters | Labeled, task-specific | Focuses on learning new parameters while keeping most of the original model fixed. |
| Soft Prompting | Few, new parameters | Labeled, task-specific | Uses learnable prompt parameters to guide the model. |
| (Continued) Pre-Training | All parameters | Unlabeled | Involves further training on large corpora to enhance model capabilities. |
Decoding
Decoding is the process of generating text from a trained model. It involves selecting words from the model’s vocabulary to form coherent and contextually appropriate responses.
Greedy Decoding
Greedy decoding picks the word with the highest probability at each step, leading to deterministic and potentially repetitive outputs.
Non-Deterministic Decoding
Non-deterministic decoding introduces randomness, allowing the model to explore various high-probability options. This can result in more diverse and creative outputs.
Temperature
Temperature is a parameter that influences the randomness of the word selection process during decoding.
- Low Temperature: Results in more predictable and focused outputs.
- High Temperature: Increases randomness, allowing for more diverse and creative responses.