This is the code for our paper "Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuning".

You can install the required libraries by running the following command:

pip install -r requirements.txt

You are also required to download the language model and put it to the folder.

You can download the data from the following link Google Drive.

Then put the data to the folder data/.

You need first to extract the function vectors from the language model. Then you can train the model with the function vectors. Finally, you can evaluate the model.

Function vector extraction This is the step to extract the function vectors for training tasks. You can run the following command:

bash scripts/fv/uni_fv.sh seq0
bash scripts/fv/uni_fv.sh seq1
bash scripts/fv/uni_fv.sh seq2

Function vector guided training Then you can train the model with the function vectors regularization term. You can run the following command:

bash scripts/train/seq0_fvg.sh {seed}
bash scripts/train/seq1_fvg.sh {seed}
bash scripts/train/seq2_fvg.sh {seed}

Model evaluation After training the model, you can evaluate the model with the following command:

bash scripts/eval/general_score.sh {exp_name}
bash scripts/eval/icl_score.sh {exp_name}
bash scripts/eval/seq0_score.sh {exp_name}
bash scripts/eval/seq1_score.sh {exp_name}
bash scripts/eval/seq2_score.sh {exp_name}

If you find this code useful, please cite our paper:

@inproceedings{jiangunlocking,
  title={Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuning},
  author={Jiang, Gangwei and JIANG, Caigao and Li, Zhaoyi and Xue, Siqiao and ZHOU, JUN and Song, Linqi and Lian, Defu and Wei, Ying},
  booktitle={The Thirteenth International Conference on Learning Representations}
}

This repo benefits from LLaMA-Factory, Function vector, and baukit. Thanks for their wonderful works.