Chinese-Llama-2-LoRA-7B

The LoRA version of Chinese-Llama-2 base on Llama-2-7b-hf.

The Github homepage is here:https://github.com/longyuewangdcu/Chinese-Llama-2/.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 2
• eval_batch_size: 1
• seed: 1
• distributed_type: multi-GPU
• num_devices: 16
• gradient_accumulation_steps: 8
• total_train_batch_size: 256
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 3.0

How to Use

Download the lora model weight in your project path

python3 inference_lora.py --model-name-or-path /llama2-7b
--lora-weights /Chinese-Llama-2-LoRA-7B/adapter_model
-t 0.7
-sa 'sample'
-i test/test_case.txt
-o test/test_case.general-task.txt

Framework versions

• Transformers 4.28.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.11.0
• Tokenizers 0.13.2