зеркало из
https://github.com/iharh/notes.git
synced 2025-10-30 21:26:09 +02:00
34 строки
2.3 KiB
Plaintext
34 строки
2.3 KiB
Plaintext
Employing advanced prompt parameters enables prompt engineers to achieve, among others, several objectives:
|
|
Control response length and stop sequence
|
|
Define the underlying model
|
|
Manage the Creativity Level
|
|
Control frequency and presence penalties
|
|
Inject start and restart text
|
|
|
|
The "temperature [0 to 1]" is a parameter that controls the creativity and randomness of the model's output.
|
|
A higher temperature (e.g., 1.0) makes the output more diverse and creative, while a lower temperature (e.g., 0.1) makes the output more focused and deterministic.
|
|
|
|
|
|
The "top_p [0-1]" (also known as nucleus sampling) dictates the scope of randomness for the language model.
|
|
It determines how many random results the model should consider based on the temperature setting.
|
|
|
|
The "stop_sequences [list of strings]" is a list of strings or tokens that, when encountered by the model, will cause it to stop generating further text.
|
|
This helps control the length and structure of the generated content, preventing the model from producing unwanted text beyond the specified stopping point.
|
|
|
|
The "frequency_penalty [-2 to 2]" parameter reduces the likelihood of the model repeating the same line verbatim by assigning a penalty to more frequent tokens.
|
|
A positive frequency penalty (e.g., 1.0) discourages the model from repeating tokens that appear frequently in the input, while a negative frequency penalty (e.g., -1.0)
|
|
encourages the model to repeat such tokens.
|
|
|
|
The "presence_penalty [-2 to 2]" parameter increases the chances of the model discussing new topics by penalizing tokens already present in the input.
|
|
A positive presence penalty (e.g., 1.0) discourages the model from using tokens already appearing in the input.
|
|
In contrast, a negative presence penalty (e.g., -1.0) encourages the model to reuse tokens from the input.
|
|
|
|
The "best_of [positive integer]" allows you to specify the number of completions (n) that the model should generate,
|
|
and it returns the best completion according to the model's internal evaluation.
|
|
This is useful when obtaining the highest quality completion from different possible results.
|
|
(n) can be an integer in the range from 1 to 20.
|
|
|
|
[
|
|
{"role": "user", "content": "Write a user story for the login process.", "settings": {"temperature": 0.8, " frequency_penalty": -1}}
|
|
]
|