This is a continuous finetuned model, based on qwen2.5-72B-Instruct, with 200,000 more instruction data.
It achieved higher performace on 13benchmark self-evaluations than the downloaded qwen2.5-72B-Instruct.
We want to validate that we can add more abilities to the model, while making all the other things better, even we don't have the original finetuing & preference datasets.