This is a modified version of the allenai/open-instruct repository used for the LLaMA-Pro project. This branch diverged from main at commit 9ebcb58. This branch implements the following features that are not present in 9ebcb58:

• Tuning specific layers while freezing the other parameters.
• Use gradient checkpointing to train.

The remaining portion of this README contains instructions to replicate training of the LLaMA-Pro model.

You can find more details in the original open-instruct.

We provide a file containing a dump of our training environment.

You can install all required packages via

pip install -r requirements.txt

You can use the following command to prepare the instruction dataset.

python open_instruct/reformat_datasets.py --dataset evol_codealpaca meta_math SlimOrca WizardLM_evol_instruct_V2_196k

Then, edit the provided .sh files to set paths based on your own system's saved locations for checkpoints and data files. The example can be found in finetune_codealpaca.

Tip: If you want to train the specific layers, you can an argument extend_layers in the script to specify the layers for training.

You can use the following command to run instruction tuning (finetuning a pretrained model to follow instructions):

bash scripts/finetune_codealpaca.sh