Fine-Tuning RoBERTa: Full Training vs Parameter-Efficient Fine-Tuning (PEFT)

Introduction

Fine-tuning pre-trained language models like RoBERTa can significantly enhance performance on various natural language processing (NLP) tasks. However, full fine-tuning of large models is resource-intensive. Parameter-Efficient Fine-Tuning (PEFT) offers an alternative by updating only a small subset of parameters, reducing computational costs while maintaining performance.

This project compares the full fine-tuning of RoBERTa with PEFT methods using LoRA (Low-Rank Adaptation). The trade-offs between these approaches are demonstrated on the AG News classification dataset.

---

Project Structure

The repository contains the following files:

1. full-finetuning.ipynb: Full fine-tuning of the RoBERTa-large model.
2. peft-roberta.ipynb: Parameter-Efficient Fine-Tuning (PEFT) using LoRA.
3. peft_vs_fulltraining.ipynb: Performance comparison of the two approaches.

---

Dataset and Libraries

Dataset

• AG News: A popular dataset for news topic classification, containing four categories: World, Sports, Business, and Sci/Tech.

Libraries

The following Python libraries are used in this project:

• transformers: Hugging Face's library for NLP models.
• datasets: Dataset handling and preprocessing.
• peft: Implementation of Parameter-Efficient Fine-Tuning techniques.
• trl: Tools for reinforcement learning with transformers.
• evaluate: Metrics computation.
• scikit-learn: Additional evaluation metrics.
• accelerate: Optimized training for distributed setups.
• bitsandbytes: 8-bit optimizers for efficient training.
• huggingface_hub: Access to Hugging Face's model hub.
• scipy: Scientific computing.
• tensorboard: Visualizing training metrics.
• matplotlib: Plotting and visualization.
• sacrebleu: BLEU score evaluation for translation tasks.

---

Notebooks Overview

1. Full Fine-Tuning (full-finetuning.ipynb)

This notebook performs full fine-tuning of the RoBERTa-large model on the AG News classification task. All model parameters are updated during training.

Training Metrics (Sample):

Step	Training Loss	Validation Loss	Accuracy	F1	Precision	Recall
500	0.315800	0.337210	0.901316	0.900811	0.905600	0.901316
5000	0.189300	0.194171	0.943158	0.943018	0.943037	0.943158
7000	0.139800	0.196201	0.949737	0.949613	0.949593	0.949737

2. Parameter-Efficient Fine-Tuning (peft-roberta.ipynb)

This notebook demonstrates PEFT using LoRA to fine-tune the RoBERTa-large model. Only a small subset of parameters is updated.

Training Metrics (Sample):

Step	Training Loss	Validation Loss	Accuracy	F1	Precision	Recall
1000	0.267400	0.321892	0.901842	0.901505	0.901446	0.901842
5000	0.236400	0.259376	0.914474	0.914257	0.914906	0.914474
7000	0.190400	0.248881	0.916053	0.915737	0.915774	0.916053

3. Comparison (peft_vs_fulltraining.ipynb)

This notebook evaluates and compares the performance of the fully fine-tuned model and the PEFT model on the test dataset.

Test Accuracy:

Model	Accuracy
Full Training	95.16%
PEFT (LoRA)	93.16%

---

Performance Comparison

Training Time and Resources

Aspect	Full Fine-Tuning	PEFT
Training Time	Longer	Faster
GPU Memory Usage	High	Low
Resource Demand	High	Suitable for limited resources

Test Accuracy

• Full Training Model: 95.16%
• PEFT Model: 93.16%

---

How to Use

1. Clone the repository:

   git clone https://github.com/ankur110/Roberta_FineTuning.git

2. Install required libraries:

   pip install transformers datasets peft trl evaluate scikit-learn accelerate bitsandbytes huggingface_hub scipy tensorboard matplotlib sacrebleu

3. Run the Jupyter notebooks in the following order:

   • full-finetuning.ipynb
   • peft-roberta.ipynb
   • peft_vs_fulltraining.ipynb

4. Customize the training parameters in the notebooks as needed.

---

Results and Insights

• Full fine-tuning achieves higher accuracy but requires significantly more computational resources.
• PEFT (LoRA) offers a cost-efficient alternative with slightly lower accuracy, making it ideal for resource-constrained environments.

---

References

• Hugging Face Transformers
• LoRA: Low-Rank Adaptation