Skip to content

Latest commit

 

History

History
136 lines (85 loc) · 4.45 KB

README.md

File metadata and controls

136 lines (85 loc) · 4.45 KB
Raw

【COLING 2025】Is Parameter Collision Hindering Continual Learning in LLMs?

Code for the N-LoRA method introduced in the COLING 2025 paper 'Is Parameter Collision Hindering Continual Learning in LLMs?', exploring novel approaches to address parameter collision issues in large language models for continual learning.
If you like our project, please give us a star ⭐ on GitHub for the latest updates.

arXiv
License Hits GitHub issues

Orthogonal and Parameter Collision

The Relationship Between Parameter Collision and Orthogonality in Continual Learning

Parameter Collision and Nuclear Norms

Parameter collision analysis and nuclear norm distributions for O-LoRA and N-LoRA. O-LoRA exhibits higher parameter collisions and nuclear norms, while N-LoRA effectively reduces collisions, ensuring better orthogonality and task-specific accuracy.

🛠️Setup

You can install the required libraries by running

pip install -r requirements.txt

You are also required to download the t5-large model from huggingface, put it to the folder named initial_model, and rename the model folder as 't5-large'.

LLaMA is also supported. You can put your llama model to the folder named initial_model and rename the model folder as 'llama'.

🚀Training and Evaluation

For t5-large:

You can reproduce our experiments of order 1 & 2 & 3 & 4 & 5 & 6 by simply running

order1:

bash scripts/order_1.sh> logs_and_outputs/order_1/logs/train_and_infer.log 2>&1 &

order2:

bash scripts/order_2.sh> logs_and_outputs/order_2/logs/train_and_infer.log 2>&1 &

order3:

bash scripts/order_3.sh> logs_and_outputs/order_3/logs/train_and_infer.log 2>&1 &

order4:

bash scripts/order_4.sh> logs_and_outputs/order_4/logs/train_and_infer.log 2>&1 &

order5:

bash scripts/order_5.sh> logs_and_outputs/order_5/logs/train_and_infer.log 2>&1 &

order6:

bash scripts/order_6.sh> logs_and_outputs/order_6/logs/train_and_infer.log 2>&1 &

The model you have trained will be saved in logs_and_outputs/order_1(2 or 3 or 4 or 5 or 6)/outputs.

The result of each task will be saved in logs_and_outputs/order_1(2 or 3 or 4 or 5 or 6)/outputs/TASK_NAME/predict_results.json.

You can also check the logs during training and infering in logs_and_outputs/order_1(2 or 3 or 4 or 5 or 6)/logs/train_and_infer.log

For LLaMA:

order1:

bash scripts_llama/order_1.sh> logs_and_outputs_llama/order_1/logs/train_and_infer.log 2>&1 &

order2:

bash scripts_llama/order_2.sh> logs_and_outputs_llama/order_2/logs/train_and_infer.log 2>&1 &

order3:

bash scripts_llama/order_3.sh> logs_and_outputs_llama/order_3/logs/train_and_infer.log 2>&1 &

🙏 Acknowledgment

This repository is adapted from O-LoRA. We sincerely thank the authors of O-LoRA for their contributions and efforts, as their work provided inspiration for this research. We also appreciate their contributions and the assistance provided by the authors.

📝 Citation

If you find this paper useful, please consider staring 🌟 this repo and citing 📑 our paper:

@article{yang2024parameter,
  title={Is Parameter Collision Hindering Continual Learning in LLMs?},
  author={Yang, Shuo and Ning, Kun-Peng and Liu, Yu-Yang and Yao, Jia-Yu and Tian, Yong-Hong and Song, Yi-Bing and Yuan, Li},
  journal={arXiv preprint arXiv:2410.10179},
  year={2024}
}