| Top K |
When appending a word to the prompt, pick out of K most likely candidates. |
| Sampling mode |
When appending a word to the prompt, pick out of most likely candidates whose total probability is reater than P. |
| Number of beams |
Track multiple copies of each prompt as it's being generated, and when done pick one with most likelihood. |
| Temperature |
When appending a word to the prompt, the greater temperature is, the more chance to pick an unlikely candidate. At 0, all generated prompts are the same. |
| Repetition penalty |
The greater the value is, the less likely repeated tearms are to appear in prompt. |
| Length preference |
Negative values tend to produce shorter prompt, positive - longer. Only works with Number of beams > 0. |
| Min length |
Minimum length of generated prompt in tokens. |
| Max length |
Maximum length of generated prompt in tokens. |