The Anima model was trained for 10,000 steps based on QLoRA's open source 33B guanaco. Training uses an H100 GPU.
This work is mainly to verify the effectiveness of the QLoRA training method, so I chose Guanaco 33B finetune training based on QLoRA. This training is more about enhancing the Chinese ability of the model. The basic logical reasoning and Knowledge capabilities of the Assume model are sufficient.
Visit Official Website