Open shamanez opened 5 months ago
No, PMC_LLaMA_7B has not undergone any instruction tunining datasets while our latest PMC_LLaMA_13B has been instruction tuned.
We have not tried LLaMA-2 for continual pre-training, since in our evaluation, LLaMA-2, compared with LLaMA, is only enhanced with instruction following ability and in basic knowledge, the gain is limited.
It sits sIs the above checkpoint the pre-trained model with unsupervised data and hasn't seen any instruction-tuning datasets?
I have another question: did you use llama2-base as the base model to conduct continual pre-training with research papers and books?