I’d like to sample from my model using a temperature schedule. For example, I’d like to start sampling with T=10, and gradually reduce it down to T=1 in first 10 tokens, and then keep sampling using T=1.
What’s the way to do that?
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Using nucleus sampling and temperature at the same time | 0 | 406 | June 27, 2023 | |
| Order of execution of Top-K, Top-P sampling along with temperature | 1 | 3416 | October 31, 2023 | |
| Greedy sampling with the new branch | 0 | 118 | July 8, 2024 | |
| In which function it is best way to use the temperature parameter .from_pretrained() or .generate() | 0 | 2241 | November 24, 2023 | |
| T5 Temperature-scaled mixing | 0 | 682 | March 18, 2022 |