ExllamaV2 version of the model created by Sao10K!

Requires ExllamaV2, which is being developed by turboderp https://github.com/turboderp/exllamav2 under an MIT license.

Main branch is 7bpw 8h

6b8h is 6bpw 8h

2.5b8h is 2.5bpw 8h

1.3, but better? I guess.

Base Merged Model ratios adjusted.

NSFL portion of Hesperus v1 dataset trained and applied.

LimaRP merged in at a ~25% weight at the end.

Subjectively better in some aspects eg. long form rp, worse than the other, eg. chat-style rps.

overall a minor improvement in my eyes.

1.5 will include Hesperus v2 dataset in its entirety.

format: alpaca.