The Phi-3-mini-4k-mango-2 is a finetune of
Phi-3-mini-4k-instruct with 4K context and 3.8B parameters.
It is a continuaton of finetuning Phi-3 (which is a great model!) to explore its properties and behaviour. More to follow.
This version of the model has had its weight layers converted to Mistral format, allowing it to run within a Mistral toolset ecosystem without change or trust_remote_code.
It seems to offer better performance than the eqivalent conversion to Llama format, which could be of interest to those using finetune toolsets yet to encompass the phi-3 model.
The process was first to convert the model weight names and config to Mistral, followed by a finetune of those weights.
