Learn the optimal parameter choice for your use case.
In most situations, the choice of parameters is guided by a single question: do I want Muse to be creative or to reply with an expected ground truth?
For detailed parameters descriptions with allowed types and values, you can visit the 📟 Developer Documentation. Here we limit ourselves to an intuitive description of the parameters that appear in the Muse Playground.
Length of the output
🧮 Number of tokens
The maximum number of tokens to generate in the completion, when you click on the
🛑 Stop Words
Sequences of characters that will cause generation to halt even before the number of tokens set has been reached. Useful if you want to stop generation at a particular point.
For example, when asking for a "positive"/"negative" classification, you might want to stop at
. or a newline, without generating anything more than "positive" and "negative".
The temperature controls the creativity of the model. For open-ended generation tasks we advise to use values around
0.9 - 1.0, for applications with a well defined answer you
should use instead values closer to zero. Alter this or Top P but not both.
🔝 Top P
Muse will consider only tokens that make up for a certain amount of probability. A value of
0.9 means that Muse will discard the least likely tokens that make up for 10% of the probability.
It is less intuitive than Temperature, and we advise to alter this or Temperature but not both.
↕️ Word Biases
Allows to control the likelihood of certain words to appear in the generation. Large negative values will ban words from appearing in the text. Large positive values will result in the exclusive generation of such words.
For example, a bias of
"hello" will ban such word from appearing in the text. A bias of
100 on the same word will make the generation output only "hello hello hello...".
😶🌫️ Presence penalty
If Muse enters in a loop here it keeps talking about the same topic over multiple sentence, you should increase this value. It penalizes tokens if they have already appeared in the text, and increases the chance of Muse talking about new arguments.
🚇 Frequency penalty
If Muse enters in a loop here it keeps repeating the same sentence verbatim, you should increase this value. It penalizes tokens depending on their frequency in the text, decreasing the chances of verbatim repetitions. It is especially useful in combination with positive word biases.