XTC
Hello all.
This model is amazing and it almost became my dayly driver. I use it for creating writing and stories. One thing I observe is that using the recommended settings, it's very good, but it still has gpt-isms (I.E. it's hard, but we will face it together, etc).
So I tried using XTC. I tried the recommended defaults of min_p=0.02, xtc_threshold=0.1, xtc_prob=0.5. It got better, but it's not enough for me.
So searching for other samplers on reddit and forums, the best combination I found are temp=0.9, min_p=0.04, xtc_threshold=0.05, xtc_prob=0.6. Now the writing style and the tone are perfect, but the model got dummer (I.E. paragraph1 describing how person A and person B never talk for a long time, but paragraph5 saying person A and person B just talked last week.)
But my guess is I enabled xtc too much.
Can anyone recommend XTC settings for this model? I am very new at LLMs and I am not sure how temp, min_p, dry, and xtc interact so it's hard to choose the best values.
Thanks!
Xtc: 0.15; 0.55
Dry: 0.85; 1.75; 2
Temp: 1
Min p: 0.01
Xtc: 0.15; 0.55
Dry: 0.85; 1.75; 2
Temp: 1
Min p: 0.01
Thanks. I'll try this out and get back to you.
I'm not a huge fan of XTC. I just tried it, and while it does add some creativity, it hampers the model heavily.
If you think about it, Mistral is plenty intelligent. Some tokens are going to be nearly assured, like, for example, names.
Consider the following test:
XTC: 0.15/0.5
Character sheet :
Heraldicus : {{char}} is a boastful adventurer. Every time he enters a taverns, he slams the doors and shouts "It is I ! Heraldicus the great !".
Chat:
{{char}}: (First message) Everything is calm in the tavern.
{{user}}: Suddenly, the doors slam loudly.
{{char}}:
Here, you'd obviously want to have the phrase "It is I ! Heraldicus the great !" somewhere in the generated response, and mistral understands that, and theses tokens will have nearly 100% probability somewhere in the sentence.
Except that with XTC, it will remove thoses tokens. I was very rarely able to get what I wanted from the instructions. Even worse, quite often, it tried to get the sentence, but it would misspell the name. Sometimes it would modify the sentence, such as "It is I ! HERALDICUS the GREAT!" (Adding caps, which I suppose it okay), but most of the time, I would get stuff like "It is I ! The great Heraldiucus !", or "Heraldinus"... You get the point.
I'm not a huge fan of XTC. I just tried it, and while it does add some creativity, it hampers the model heavily.
If you think about it, Mistral is plenty intelligent. Some tokens are going to be nearly assured, like, for example, names.Consider the following test:
XTC: 0.15/0.5Character sheet :
Heraldicus : {{char}} is a boastful adventurer. Every time he enters a taverns, he slams the doors and shouts "It is I ! Heraldicus the great !".Chat:
{{char}}: (First message) Everything is calm in the tavern.
{{user}}: Suddenly, the doors slam loudly.
{{char}}:Here, you'd obviously want to have the phrase "It is I ! Heraldicus the great !" somewhere in the generated response, and mistral understands that, and theses tokens will have nearly 100% probability somewhere in the sentence.
Except that with XTC, it will remove thoses tokens. I was very rarely able to get what I wanted from the instructions. Even worse, quite often, it tried to get the sentence, but it would misspell the name. Sometimes it would modify the sentence, such as "It is I ! HERALDICUS the GREAT!" (Adding caps, which I suppose it okay), but most of the time, I would get stuff like "It is I ! The great Heraldiucus !", or "Heraldinus"... You get the point.
Yep that's my observation as well regarding XTC. The writing style changed, but the model got dumber. I'm using it because I'm trying to eliminate the model's gpt-isms (I.E. unbreakable bond, always there forever, that kind of stuff), what do you suggest I can try? Alsoany recommendations for DRY?