🧠 MAP – Student Misconception Detection (DeBERTa-v3)

Transformer-Based NLP Classification | Kaggle Competition Project

This project tackles the MAP – Charting Student Math Misunderstandings Kaggle competition.

The goal is to classify mathematical misconceptions from open-ended student explanations using a transformer-based deep learning model.

---

🎯 Problem Overview

Students explain their reasoning in free-text format.
The task is to predict:

Category : Misconception

based on:

• Question text
• Multiple-choice answer
• Student explanation

This is a multi-class NLP classification problem with strong class imbalance.

---

🏗 Model Architecture

Input Text
(Question + MC Answer + Student Explanation)

→ Tokenization (DeBERTa-v3)
→ Transformer Encoder
→ Classification Head
→ Softmax
→ Category:Misconception Prediction

---

⚙️ Implementation Highlights

• ✅ DeBERTa-v3 Backbone
• ✅ Stratified K-Fold Cross Validation (3 Folds)
• ✅ Mixed Precision Training (AMP)
• ✅ Cosine Learning Rate Scheduler
• ✅ AdamW Optimizer
• ✅ Layer Freezing for Faster Training
• ✅ Fold Ensembling for Final Submission
• ✅ Macro F1 Evaluation

---

📊 Evaluation

Primary Metric: Macro F1 Score

• Stratified splitting ensures label balance
• Validation performance tracked per fold
• Best fold model saved and used for ensembling

---

🛠 Tech Stack

• PyTorch
• Hugging Face Transformers
• Scikit-learn
• Pandas / NumPy
• Google Colab (GPU T4)

---

📸 Notebook

Colab Notebook: (https://colab.research.google.com/drive/1p7GqShMU9kcon3isXY7xAhMCfBCfrcqu?usp=sharing)

---

🧠 Engineering Learnings

• Handling extreme class imbalance in NLP classification
• Implementing stratified K-Fold validation
• Managing transformer training stability
• Debugging mixed precision & gradient instability
• Designing clean Kaggle submission pipelines

---

🔮 Future Improvements

• Use DeBERTa-v3-Large
• Apply label smoothing
• Try class-balanced loss
• Apply pseudo-labeling
• Add model distillation for faster inference