explanations
explanations
Please, I would like to know how it has been improved from the original model and how
Thank you 😊
It was rebuilt from source into float32, converted the ggufs at float 32, added additional parts of source to the ggufs in float 32 (all ggufs , no compression) and then a stable imatrix dataset used to finalize everything.
What do you mean by: "added additional parts of source to the ggufs in float 32"?
I tried the model, but this one compared to the original, falsifies the history data a lot. it is not reliable
Thank you for your feedback.
This have been investigated, and issue(s) located.
New quants are being built and will be uploaded shortly.
This should improve overall performance at all quant levels above current levels.
missed opportunity to name it Fimbulvetr-11B-Plus-Ultra for the MHA fans.