Fine-tuning GPT-2 using LORA to infer Emotion class of Twitter messages

Fine-tuning the GPT-2 model to classify emotions in tweets. The process includes loading a pre-trained foundation model, performing parameter-efficient fine-tuning (PEFT) using LoRA, evaluating performance, and comparing the foundation model with the fine-tuned model.

Choices

• Model:
  • GPT-2 is used because it is compatible with the sequence classification task and compatible with LoRA.
• PEFT Technique:
  • LoRA (Low-Rank Adaptation) is utilized as it allows for efficient fine-tuning without significant impact on the original model weights.
• Evaluation:
  • Hugging Face's Trainer .evaluate method is used to compare the performance of both the foundation and the fine-tuned models.
• Dataset:
  • The dataset consists of labeled tweets on emotion, provided by Hugging Face datasets.
    • Dataset Link

Steps

1. Choose the dataset: DONE
2. Choose the foundation model: DONE
3. Perform inference for the text classification task with the foundation model: DONE
4. Evaluate the performance of the foundation model: DONE
5. Load the foundation model as a PEFT model: DONE
6. Define the PEFT/LORA configuration: DONE
7. Train the LoRA model with Hugging Face Trainer: DONE
8. Evaluate the PEFT model: DONE
9. Save the PEFT model: DONE
10. Load the saved PEFT model from local storage: DONE
11. Run inference and generate text/label with the tuned model: DONE

Results

• Can be found in the notebook run
• Evaluation accuracy of the foundation model on this task is 'eval_accuracy': 0.096
• While the Evaluation accuracy of the tuned model is 'eval_accuracy': 0.9225
• This is almost a 10x increase