Skip to main content

🟢 LLM Settings

ILikeTheItWhenThatTrainsFrogsProbabilitiesof next token0.050.30.30.150.150.05

takeaways
  • Understand Temperature, Top P, and Maximum Length

Introduction

We can use certain LLMa 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:
Astronaut

Prompt


What are 10 weird, unique, and fun things to do at the beach? Make a list without descriptions.
Robot

ChatGPT Output (Temperature=0, Top P = 1)


1. Build a sandcastle
2. Collect seashells
3. Play beach volleyball
4. Fly a kite
5. Have a picnic
6. Try paddleboarding
7. Play frisbee
8. Go snorkeling
9. Take a long walk along the shore
10. Watch the sunset
Robot

ChatGPT Output (Temperature=1.5, Top P = 1)


1. Wade in the shallow waters blowing bubblegum bubbles
2. Create intricate sandcastle sculptures
3. Join an improvised beach volleyball match
4. Take a scenic hike along the cliffs nearby
5. People-watch and guess each person's beach-related story
6. Collect seashells to create your own piece of beach art
7. Learn to surf or try other water sports
8. Kickstart a spontaneous sand fight
9. Attempt shoreline fishing like the locals
10. Engage in friendly competition by organizing a sandcastle building contest

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.

note

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 Pb 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 is 50%.
  • Then adding roof -> total becomes 75%.
  • Next comes wall, and now our sum reaches 90%.

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. It also helps control cost, as the length is shared between the input in the Playground box and the generated response.

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 after the closing salutation, which keeps the email short and to the point.

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 ocurred 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


  1. A more technical word is "configuration hyperparameters"↩
  2. Also known as Nucleus Sampling↩