How Do I Use Max Tokens, Temperature and Other Model Settings?
In the Bulk Publishing AI plugin, you can adjust the model settings for each variable.
According to ChatGPT, these settings exist because they provide users with a way to control and fine-tune the AI’s output according to their specific needs and preferences.
Here is what each of them does:
Model:
This refers to the specific version of the AI model being used. Different models have varying sizes, training data, and capabilities. A larger model, like GPT-4, is more powerful and can generate more accurate and coherent responses than smaller models, but it may also be slower and more resource-intensive.
Max tokens:
This setting controls the maximum length of the generated response. A token can be as short as one character or as long as one word. Higher values allow for longer responses, while lower values limit the response length. If you set max tokens to a low number, the output might be cut off and not make sense.
Temperature:
This setting determines how creative or random the generated text will be. Higher values (e.g., 1.0) result in more creative and diverse responses, while lower values (e.g., 0.1) make the output more focused and deterministic, sticking closely to the input text and choosing more common words.
Top P:
Top P is a way to control the randomness of the AI’s output by only considering the most probable words for the response. Imagine you have a bag of differently colored balls, where each color represents a word. The probability of each word is like the number of balls of that color in the bag.
When the AI generates a response, it picks a ball (word) from the bag. With Top P, you set a threshold (like 90%) to include only the most probable words (colors) that together make up 90% of the total probability.
In other words, Top P removes the least probable words from the bag, making the AI choose from a smaller set of more probable words. This makes the output more focused and coherent.
Higher Top P values (closer to 1) include more words (colors) in the bag, resulting in more diverse responses. Lower Top P values (closer to 0) include fewer words, making the output more similar to the input text and less random.
Frequency penalty:
This setting helps control the repetition of words and phrases in the generated text. Higher values (e.g., 1.0) penalize frequently occurring words, making the model less likely to repeat them, resulting in more diverse output. Lower values (e.g., 0.0) don’t penalize frequent words, which may lead to more repetitive output.
Presence penalty:
This setting is similar to frequency penalty but focuses on penalizing words that have already appeared in the generated text. Higher values (e.g., 1.0) encourage the model to use a wider variety of words, while lower values (e.g., 0.0) allow for more repetition of words already present in the text.