Request to delete corrupt models
Hi,
I hope all is well. I hate to bother you but I need to request that you delete a few models.
it seems that I had an issue with a new process I was using that caused some models I merged last night to become corrupt.
it turns out they have a tendecy to return gibberish. not sure what the issue is yet but it's 100% on my end.
the models are:
mradermacher/Eros_Scribe-9b-GGUF
mradermacher/NeuralStar_FusionWriter_13b-GGUF
mradermacher/NeuralStar_Fusion-9B-GGUF
The other models are awesome and I really appreciated the f16 and the iQ quants!!
No rush just whenever you can get around to it
thanks,
Mike
Let me a) delete these and b) convince you otherwise, at least in some cases: I checked a model, and it practically always generates gibberish, and, yeah, it's a service to delete those. However, form your description ("tendency") it sounded as if these models do work, but sometimes fail. In that case, I'd ask you to reconsider - it's great that you have high quality standards, but in most cases, IMnsHO, the best way to proceed is to update the model card saying sth. like "this model is very bad/broken/whatever". Yes, it might be embarassing bad quality to you, but there are lots of models where the creator is very unhappy, but some people really love them.
Anyway, if you come up with a model you really like and I overlook it somehow, don't hesitate to point me at it :) Cheers!
yes, you'r right. In hindsight, I wish I would have left one of them up. Then at least, I could have reached out to the community. I initially though it was oobabooga(my first 4x7b base model responds in russian half the time in ooba but is fine w gguf?) but when the quants came out I realized they were hosed.
The other thing is that since the whole Lama 3 thing it seems like people want to blame the quants and I didnt want you to have to field those complaints...
Anyway, thanks again
Yeah, llama-3 had a really bad start. And it seems really hard to finetune it, as well. But I am sure things will improve :)