You have my gratitude
The model merges around here were beginning to look like this:
I appreciate your effort. No joke. Gold star or something. Platinum, even.
Thank you for your praise. If you have any questions, please give us feedback so that I can iterate on it in subsequent versions of training.
I will definitely let you know. What I would really like to have, though, is a version that uses the llama tokenizer. KoboldAI doesn't support --trust-remote-code, and getting it to work is a huge pain (it's more complex than just adding the argument to several function calls). I think Yi updated their model to not require that anymore, as my impression is that there really isn't any reason the model needed to do that to begin with. They just created it with just enough of a breaking change to require it, though I don't recall the specifics. Actually, you're probably aware of it and understand it a lot more than I do, haha.
Hi. I really like the concept, but it seems somewhat "overtrained". Perhaps I'm not prompting it right, or perhaps my settings are bad? If you want, I could find you more data to train it with. It was a while ago, so I can't remember off the top of my head, but I have found datasets used in psychological studies. They scraped NSFW chat data for one of them. There is a lot of raw data like this online. There are sites which catalog it it as a corpus for use in linguistics, etc.
Hi. I really like the concept, but it seems somewhat "overtrained". Perhaps I'm not prompting it right, or perhaps my settings are bad? If you want, I could find you more data to train it with. It was a while ago, so I can't remember off the top of my head, but I have found datasets used in psychological studies. They scraped NSFW chat data for one of them. There is a lot of raw data like this online. There are sites which catalog it it as a corpus for use in linguistics, etc.
Sorry just saw this message, I would be happy to use more data to train version 2
Just saw your reply, lol. I'll try to track that stuff down again.