🧠 RoBERTa Token Classification with Additional PLODv2 Data

This repository contains two Python scripts that fine-tune a roberta-base model for Named Entity Recognition (NER) on the PLOD-CW-25 dataset, with 25% and 50% additional samples from the PLODv2-filtered dataset. The goal is to evaluate performance gains in token-level classification using partial data augmentation.

---

📁 Files

Filename	Description
RoBERTa+25%data.py	Fine-tunes roberta-base on PLOD-CW-25 with 25% of PLODv2-filtered samples added to training and validation sets.
RoBERTa+50%data.py	Fine-tunes roberta-base on PLOD-CW-25 with 50% of PLODv2-filtered samples added to training and validation sets.

---

🔍 Dataset Overview

• PLOD-CW-25: Legal NER dataset with annotated tokens from case law.
• PLODv2-filtered: A filtered version of PLODv2 for optional fine-tuning enhancement.

Each script:

• Loads both datasets.
• Randomly samples a fraction of PLODv2-filtered.
• Merges it with the original training and validation splits.
• Converts the data into Hugging Face Datasets format.

---

🔧 Model Architecture

• Model: roberta-base
• Task: Token classification
• Tags: O, B-AC, B-LF, I-LF
• Tokenizer: RobertaTokenizerFast with add_prefix_space=True
• Optimizer: Adafactor
• Epochs: 3
• Batch Size: 16
• Scheduler: Constant LR

---

📊 Metrics & Evaluation

• Evaluation Metric: seqeval
• Reports:
  • Overall: Precision, Recall, F1, Accuracy
  • Entity-wise: Per-class F1
  • Visuals: Confusion Matrix & Bar Plot for metrics

---

📈 Visual Outputs

Each script generates:

• 📋 Classification report
• 📊 Bar chart for precision/recall/F1 by entity
• 🔲 Confusion matrix for true vs. predicted labels

---

📊 Side-by-Side Result Comparison

Metric	25% Additional Data	50% Additional Data
Precision	~0.88	~0.90
Recall	~0.89	~0.91
F1 Score	~0.88–0.89	~0.90+
Accuracy	~0.89	~0.90

---

🛠️ Installation

pip install datasets transformers huggingface_hub evaluate seqeval nbconvert

---

📌 Notes

• Scripts are optimized for execution in Google Colab.
• Designed for experimentation with data augmentation in NER tasks.
• Results can help in assessing trade-offs between more data vs. training time.

---

📤 Output Example

{
  "precision": 0.89,
  "recall": 0.91,
  "f1": 0.90,
  "accuracy": 0.90
}

📁 Check the detailed output in the confusion matrix and classification report printed at the end of each script.

---

✍️ Author

Aaditya Singh – MSc Data Science
For academic use and performance benchmarking of NLP models in the legal domain.

---