Notable Improvements?
Ive been using Magnum12b for a while before discovering yours here. could you give me a bit of insight as to what exactly are the pros and cons between this and magnum? magnum itself is in this merge, so would you say this is any improvement? there isn't much info as to what the merge is capable of sorry im just curious.
Tbh I personally found that merging Magnum with Celeste consistently makes dialogue more believable and lifelike (due to half of Celeste's dataset being natural data). Also it seems to somewhat curb both Magnum's and Celeste's NSFW and violence biases.
Magnum isn't a merge btw, idk where you heard that, it's a full finetune.
interesting to know thank you for the info.
Magnum isn't a merge btw, idk where you heard that, it's a full finetune.
i never said it was a merge i said it was IN this merge. i was saying because i was blown away with the Magnum finetune and was really really curious if it could get "better" seeing as it was used as an ingredient here.
i never said it was a merge i said it was IN this merge. i was saying because i was blown away with the Magnum finetune and was really really curious if it could get "better" seeing as it was used as an ingredient here.
whoops, sorry. Seems like I need to sleep more :)
Model is amazing! Good job, but does it still use ChatML properly? I notice the LLM ends every response reliably with '<|im_' which is a little weird, regardless if I use ChatML or Mistral instruct formats. (It's not being cut off either, it happens in 600 token responses and my limit is set to 900 tokens)
Hm, it should work properly with ChatML. Both merged checkpoints have ChatML embeds trained (and in the same way), and correct vocab was imported. Check if your backend sets EOS token correctly