Main settings
Response Length (tokens)
366
Context Size (tokens)
4096
☐ Unlocked
Temperature
0.70
(Though you could go up to 0.98
without much trouble)
Rep. Pen.
1.04
(Can do down to 1.01
if you're not a fan of the output length)
Rep. Pen. Range
0.00
Encoder Rep. Pen.
1.00
No Repeat Ngram Size
0.00
Min Length
0.00
Top K
0.00
Top P
0.90
Typical P
1.00
Top A
0.00
Tail Free Sampling
1.00
Epsilon Cutoff
0.00
Eta Cutoff
0.00
☑ Streaming (Only if you actually plan to use Remote Moe's ssh nonsense to proxy over port 5005)
☑ Do Sample
☑ Add BOS Token
☐ Ban EOS Token
☑ Skip Special Tokens (Sometimes weird characters show up if this isn't toggled, but it doesn't prevent ***
from randomly appearing.)
CFG Scale
1.00
Number of Beams
1.00
Length Penalty
1.00
☐ Early Stopping
Penalty Alpha
0.00
Mirostat Mode
2.00
Mirostat Tau
5.00
Mirostat Eta
0.10
Seed
-1
Advanced Formatting
Context Template
Custom Stopping Strings (KoboldAI/TextGen/NovelAI)
["\nUser:", "\n{{user}}:", "\nYou:", "\n\n***", "\n***", "\n***\n", "\n\n**Note: ", "\n\nNote: ", "\nSuggestion:", "# Suggestion:", "\n**Storyline:**", "# **Storyline:**", "# Storyline:", "\nStoryline:", "**Note:", "</s>", "\n___", "#### Narration:", "# Narration:", "## Narration:", "### Narration:", "\n# Takeaway:", "#####", "Me: ", "\n[(Human)", "\nNOTE:", "\n#", "<END>", "\n—Game session ends here—", "\nIB:", "|user|>", "user|>"]
(^this is to deal with a whole bunch of annoying edge cases where the model acts goofy)
System Prompt (New)
(^For some reason, this above prompt makes everything act a shit-ton more like C.AI, but in a good way.)
Input Sequence
Output Sequence
System Sequence Prefix
Stop Sequence
Instruct mode
Presets Metharme
☑ Enabled
☑ Wrap Sequences with Newline
☑ Replace Macro in Sequences
☑ Include Names
☑ Force for Groups and Personas
Token Padding
0