Context
Tools relative to Context. Explained on the corresponding page.
Phrase Bias
Phrase Bias allows you to tell the AI to increase or decrease the probability of single or groups of tokens. You may have 1024 biases enabled at once. Any extra will cause generation to fail.
Rather than ban tokens outright, this is more of a direction to produce specific output, or avoid specific outputs without fully eliminating them.
Keep in mind that Phrase Bias is case sensitive. However, a variation of the first token with a space will be considered, even if you do not specify it. (This does not apply if the tokens are ;, :, <, >, &, @, #, %, ^, \n, and the entire range of unicode characters from u3000-u9faf and uff00-uff9f.)
Bias is collected in Groups with a certain Bias Level, which is applied to all phrases and tokens put in the Group.
To create a new Bias Group, click the ➕ button. The currently selected group can be deleted by clicking 🗑️.
To enter a new phrase or token in the Group, enter it in the text box and press Enter.
If you wish to insert token IDs (such as [198] for newlines), encase their numerical ID in Square Brackets: []
If you wish to insert the exact input, as is, case sensitive and with exact spacing, encase in Curly Braces {}.
Bias Level is a non-linear Scale from -2 (less likely) to +2 (more likely). It is not an arithmetic scale: -2 is much stronger than simply "twice less likely than -1". The opposite is true for +1 and +2. Generally, a bias of -0.1 to +0.1, usually lower, works well for most uses cases.
Ensure Completion After Start
If checked, the AI will make sure Phrases are completed if their first token is generated.
For example, if you've given a positive bias to the phrase "blue business suit" with this function turned off, the AI might start generating outputs featuring blue lights, blue planets, or people with blue eyes. With Ensure Completion After Start turned on, the word "blue" will always be followed by the phrase "business suit".
Unbias after Generation
If checked, the AI will disable the bias for the remainder of the generation once it has been applied once. Following the previous example, the first instance of "blue" will output "blue business suit", but all following instances will work as normal.
Ban Token
Any Tokens added here will have their likelihoods reduced to zero. This means they will not appear in Generations. As this adjusts the relationships between Tokens, this will have an impact on the phrasing chosen by the AI. Be careful about what you ban, because this can heavily disrupt output if used incorrectly.
When you add a new token to the banlist, it also adds any case-sensitive variation of the token, and the token with a preceding space as well.
To prevent that from happening, and exclusively ban this token, addcurly braces { } around the token you want to ban before pressing enter. You may have 2048 bans at once, any extra will cause generation to fail.
Banned Tokens
Relating to the previous setting, this field shows every Token currently blacklisted for generation. Clicking one of these tags will remove it from the list.
If you inserted an exclusive token (using curly braces), it will be displayed with brackets around it.
Ban Bracket Generation
At times, you may wish to include hints to the AI that are not considered for text generation. These can be encapsulated in square brackets ([ and ]) to relay information that will affect the Current Context while not being considered part of the actual text.
These most often take the form of hints. For more information on what to put between square brackets, see [Keeping Track](Advanced Writing#Keeping Track).
Stop Sequence
If you wish to end your Generations upon reaching a specific Token, simply add it to this field. Doing so will cause the Generation to end prematurely upon generating the Token.
This can be used to trim the output to single sentences by inputting punctuation, or to increase the accuracy of Lorebook entries by pausing Generation when a defined word is reached.
Min Output Length
This adjusts the minimum number of Tokens returned by the AI before having a chance to generate the Stop token. It will never be below this target, but may exceed it.