Manual and automatic parameters in AI training and output

Language Models, such as GPT-3, have billions or trillions of parameters that allow them to understand and generate human-like text. These parameters are the numerical weights that the model uses to learn and represent language patterns during training. While some parameters, like the temperature in a generative model, are set explicitly for controlling behavior, the vast majority of parameters are learned automatically through a process called training, and humans do not set them manually. You can read more about the other manually-configured parameters here: https://michaelehab.medium.com/the-secrets-of-large-language-models-parameters-how-they-affect-the-quality-diversity-and-32eb8643e631

Manual Parameters:

Adapted from the blog above, we see some manual parameters are:  Some of the common LLM parameters are temperature, number of tokens, top-p, presence penalty, and frequency penalty.

• Temperature: Temperature is a hyperparameter used in generative language models. It controls the randomness of the model's output. Higher values (e.g., 1.0) make the output more random, while lower values (e.g., 0.2) make it more deterministic and focused.
• Number of Tokens: This parameter specifies the maximum length or number of tokens in the generated text. You can set it to limit the length of the generated content, which is useful to prevent overly long outputs in tasks like text generation and summarization.
• Top-p (Nucleus Sampling): Top-p, also known as nucleus sampling, is a parameter that influences the diversity of generated text. It sets a probability threshold, and the model selects from the most probable tokens until the cumulative probability exceeds this threshold. It helps control the output's diversity and prevents excessively repetitive text.
• Presence Penalty: Presence penalty is a parameter used in text generation. It encourages the model to avoid repeating the same tokens too frequently in its output. This is useful for making the generated content more diverse and less repetitive.
• Frequency Penalty: Frequency penalty is another parameter used for text generation. It discourages the model from using very common words or tokens too often in the generated text. This can help make the output more varied and avoid overusing generic terms.

Here's how it works:

• Training Data: Language models are trained on a massive dataset containing text from the internet, books, articles, and various sources. This data is used to teach the model about the structure and patterns of human language.
• Neural Network Architecture: The architecture of the model is designed to capture these language patterns. In the case of GPT-3, it uses a deep neural network with many layers, and each layer contains a large number of parameters.
• Learning via Optimization: During training, the model adjusts these parameters to minimize the difference between the text it generates and the text in its training data. It uses a process called stochastic gradient descent (SGD) or variants of it to adjust these parameters. Stochastic Gradient Descent (SGD) is a popular algorithm used in machine learning for training artificial neural networks. It is an optimization algorithm that helps to minimize the cost function of a model by updating its parameters based on the gradient of the error with respect to the parameters. In simpler terms, it helps to adjust the weights and biases of the neural network during training to improve its performance. SGD is commonly used in deep learning applications such as image recognition, natural language processing, and speech recognition.
• Fine-Tuning: After the initial training on a large dataset, the model can be fine-tuned on specific tasks or datasets. In this step, a smaller set of parameters may be adjusted to adapt the model to a particular task, such as translation, summarization, or question-answering.
• Inference: Once the model is trained, it can be used for various natural language understanding and generation tasks. The billions or trillions of parameters it has learned are essential for it to generalize and perform well on a wide range of tasks.

Setting the individual parameters manually is not practical or feasible because there are too many of them. Instead, the model's parameters are optimized during training to capture the statistical regularities in the training data. This process allows the model to learn and generate text that appears human-like, even though no human has explicitly set each parameter. The massive number of parameters in these models is what makes them so powerful and capable of understanding and generating natural language text effectively.

This explanation mostly generated by ChatGPT and MiniOrca.