Positivity Bias?
I'm using a text completion preset in ST with temp: 1.05, MinP: 0.16, PPR: 2048, DyTemp: 0.8-1.25, SO:6 0 1 3 4 2 5, and ChatML. So, I started a NSFW group chat card that had me worried to begin with due to 'positivism', though in fact the card itself hints at the 'sanitized' and 'corporate' lifestyles of two of the characters in the group. However, they did the deeds. It was a funny and refreshing twist. I'm not sure how much my settings affected the output, but they are the same tentative settings I'm using for yi-34b-200k-dare-megamerge-v8. All in all, it's a smart model as far as I've tested so far. It was refreshing to see very little GPTism until very late in the chat when it said, "He knew then that there was no turning back." XD
:) glad to hear it, thanks for testing. Ill definitely add 'no turning back' as a bad phrase in the next iteration of the model.
I have one other observation. I'm running my models from old spinning metal (HD), and nothing loads faster than this one in the 13b+ range. It's doing a better job at converting a role play chat into a short story than any smaller models thanks to its much larger context and intelligence.
"Together, they would face these trials head-on, united by an unshakable bond that transcended words and blossomed into something far greater than either could have ever imagined."
Kill me now!! XD
I find myself smiling at the irony sometimes.
omg, thats GPT in all its 'glory' lol, i've masured probablitity to produce exactly this sequence, its below 1% but it can happen.