Note: the vaner is based on LLama-2-7b-hf, you need to download this base model.
Mix multiple datasets for Instruction-Tuning with DBR
CUDA_VISIBLE_DEVICES=0 python unllama_train_vaner.py unidev_kgmix2 128 mt vaner
Single datasets Instruction-Tuning with DBR
CUDA_VISIBLE_DEVICES=0 python unllama_train_vaner.py ncbi 128 mt vaner
If you lack computational resources for training, you can also download our pre-trained models. We provide two models:
bjy_unidev_kgmmix2:
This model has been fine-tuned with a mix of 8 datasets and has undergone DBR.
bjy_unidev_kgmmix2_nodataname:
This model has also been fine-tuned with a mix of 8 datasets, but does not specify the name of the dataset in the construction of the prompt, corresponding to the VANER_adapt method in the paper.
Change 'model_id' in unllama_eval_vaner.py as the path of LLama-2-7b-hf.
Evaluation script
CUDA_VISIBLE_DEVICES=0 python unllama_eval_vaner.py ncbi vaner_unidev_kgmix2_checkpoint-number