Config.json seems to be from one of your older models!
You better use the correct one, otherwise your quants might turn out bad, as they might use stuff from in there. Also some GUI frontends pull data from huggingface when downloading a model for the default settings of the model.
They are all the same configs as this file has not really changed. thank you for pointing that out though! I was away that night and uploaded that config file from my laptop
Np. I cross-checked with phi-2's original config and it mostly overlaps. There is only one change that made me think you reused the config from Qwen2, Yi-1.5 or another model, which was the "max_position_embeddings": 32768, in the config.json. I suppose you used a custom value, because the base model is only trained on 2048, but your finetuning dataset holds roughly up to 8k tokens of context.