tags:
- roleplay
I removed previous version of the presets as I feel they were bad.
Thanks to shrinkedd for some ideas.
SillyTavern Presets
Usage
Make sure to grab both context and instruct templates. It should look something like this.
When using these presets you must set Example Messages Behavior: Never include examples otherwise they will be sent twice.
The reason for this, is because I explicitly set for them to be sent. The default behavior is for them to just be added at the end of the context prompt.
Important: Edit the system prompt with your preferred formatting if you don't like action/narration "dialogue".
Character Cards
When using these presets it is recommended to use Personality Summary, Scenario, Example Dialogue.
Example Dialogue should look like this (Ali-Chat works as well.):
It may look like this:
{{char}}: *looking at her book* "You can tell a lot about a person by what they choose to read."
Or this:
{{char}}: "You can tell a lot about a person by what they choose to read."
Samplers
I have decided to remove old samplers and only keep basic presets, I want people to play around and find what works best for them. Change context to desired context length
[Dynamic-Temp]Roleplay has a tendency to break asterisk formatting, however is also stops alot of the GPTism.
Temperature Feel free to play with this one, lower values are more grounded.
Min-P Higher values chop off more probabilities.
Values between 0.025 - 0.10 are good, personally I would use 0.075 or lower.
Repetition Penalty Tries to decrease repetition.
Do not set it higher than 1.2.
1.05 - 1.15 seem to work fine.
Rep Pen Range The range of tokens which Repetition Penalty can see. I have it set to 2048.
Frequency Penalty Decreases repetition.
Presence Penalty Increases word variety.
Dynamic Temperature Min and Max temps, free to change as desired.
Exponent, do not set Exponent higher than the default of 1.
You might want to try playing around and setting it lower than 1, this pushes lower probabilies higher.
When setting exponent lower than 1, set Min-P a little higher (0.075)
Smooth Sampling This one is great, smoothens out probabilities.
Lower is more diverse.
Recommended range 0.1 - 0.3