Learn PromptingLLM Settings
Last updated on August 7, 2024
Introduction
We can use certain LLM settings to control various aspects of the model, such as how 'random' it is. These settings can be adjusted to produce more creative, diverse, and interesting output. The Temperature, Top P and Max Length settings are most important, but we describe every setting that the OpenAI Playground allows you to modify.
Temperature
Temperature regulates the unpredictability of a language model's output. With higher temperature settings, outputs become more creative and less predictable as it amplifies the likelihood of less probable tokens while reducing that for more probable ones. Conversely, lower temperatures yield more conservative and predictable results. The following example illustrates these differences in output:
Prompt
ChatGPT Output (Temperature=0, Top P = 1)
ChatGPT Output (Temperature=1.5, Top P = 1)
The output produced with a higher temperature setting offers a more imaginative and diverse list of activities to do at the beach. This can be very useful for creative writing.
If you adjust the temperature too high, you can get non-sensical outputs like
Start a sponge-ball baseball home run contest near Becksmith Stein Man Beach.
Top P
Top P is a setting in language models that helps manage the randomness of their output. It works by establishing a probability threshold and then selecting tokens whose combined likelihood surpasses this limit.
For instance, let's consider an example where the model predicts the next
word in The cat climbed up the ___. The top five words it might
be considering could be tree (probability 0.5),
roof (probability 0.25), wall (probability 0.15),
window (probability .07) and carpet, with
probability of .03.
If we set Top P to .90, the AI will only consider those tokens which cumulatively add up to at least ~90%. In our case:
- Adding
tree-> total so far is50%. - Then adding
roof-> total becomes75%. - Next comes
wall, and now our sum reaches90%.
So, for generating output, the AI will randomly pick one among these three options (tree, roof, and wall) as they make up around ~90 percent of all likelihoods. This method can produce more diverse outputs than traditional methods that sample from the entire vocabulary indiscriminately because it narrows down choices based on cumulative probabilities rather than individual token
Maximum Length
The maximum length is the total # of tokens the AI is allowed to generate.
This setting is useful since it allows users to manage the length of the
model's response, preventing overly long or irrelevant responses. The length
is shared between the USER input in the Playground box and the
ASSISTANT generated response. Notice how with a limit of 256
tokens, our PirateGPT from earlier is forced to cut its story short
mid-sentence.
This also helps control cost if you're paying for use of the model through the API rather than using the Playground.
Other LLM Settings
There many other settings that can affect language model output, such as stop sequences, and frequency and presence penalties.
Stop Sequences
Stop sequences tell the model when to cease output generation, which allows you to control content length and structure. If you are prompting the AI to write an email, setting "Best regards," or "Sincerely," as the stop sequence ensures the model stops before the closing salutation, which keeps the email short and to the point. Stop sequences are useful for output that you expect to come out in a structured format such as an email, a numbered list, or dialogue.
Frequency Penalty
Frequency penalty is a setting that discourages repetition in the generated text by penalizing tokens proportionally to how frequently they appear. The more often a token is used in the text, the less likely the AI is to use it again.
Presence Penalty
The presence penalty is similar to the frequency penalty, but flatly penalizes tokens based on if they have occurred or not, instead of proportionally.
Determinism Note
Even when Temperature and Top-P are set completely to zero, the AI may not give the same exact output every time. This is due to randomness in GPU (graphics processing unit) calculations being done in the AI's "brain".
Conclusion
In conclusion, mastering settings like temperature, top p, maximum length and others are essential when working with language models. These parameters allow for precise control of the model's output to cater to specific tasks or applications. They manage aspects such as randomness in responses, response length and repetition frequency among other things—all contributing towards improving your interaction with the AI.
Partly written by jackdickens382 and evintunador