I'm applying this to the Ghost 8B Beta (128k) chat version online here and it seems to work.
In general, I have not yet fine-tuned and tested the parameters against the original model (even the current version is online) but I have actually noticed that the context is long but still ensures very good quality, for example here.
This is a quick issue to share this joy with your research team. Thank you very much~