jatinmehra/Qwen-3-8B-Instruct-MATH-Misconception-Annotation-Project

Qwen3-8B Math Misconception Classifier

Model Description

This model is a fine-tuned version of Qwen/Qwen3-8B for identifying and classifying student mathematical misconceptions. The model analyzes student explanations of math problems and predicts the specific misconception category they exhibit.

Model Architecture: Qwen3-8B (8 billion parameters)
Task: Multi-class Text Classification (65 misconception classes)
Performance: MAP@3 Score of 0.944

Intended Use

Primary Use Cases

• Identifying mathematical misconceptions from student explanations
• Educational assessment and personalized learning
• Automated feedback systems for math education
• Research in mathematics education

Out-of-Scope Use

• General text classification tasks outside of math education
• Real-time production systems without human oversight
• Any application where misclassification could lead to harm

Training Details

Training Data

The model was trained on the MAP Charting Student Math Misunderstandings dataset, which includes:

• Mathematical questions with multiple choice answers
• Student explanations for their answer choices
• Labels indicating whether the answer was correct
• Misconception categories and specific misconceptions

Training Procedure

Input Format:

Question: {QuestionText}
Answer: {MC_Answer}
Is Correct Answer: {Yes/No}
Student Explanation: {StudentExplanation}

This structure provides the model with full context: the question, the student's answer choice, whether it's correct, and their reasoning.

Preprocessing Steps:

1. Created target labels by combining Category and Misconception columns
2. Transformed labels into numerical format using label encoding
3. Identified correct answers and merged this information into the training data

Training Configuration:

• Model: Qwen 3 8B
• Method: Full Fine-tuning
• Learning Rate: 2e-5
• Epochs: 3
• Batch Size: 16
• Precision: Mixed precision (FP16/BF16)

Model Evaluation

The model was evaluated using the MAP@3 metric on the validation set from the competition, achieving a score of 0.944.

Evaluation Procedure:

• Predictions were generated for the validation set
• MAP@3 score was calculated based on the competition's evaluation script

Limitations & Bias

• The model is specifically tuned for the MAP competition dataset and may not generalize to other text classification tasks.
• There may be biases present in the training data that could affect the model's predictions.
• Misclassifications could occur, especially in cases of ambiguous or unclear student explanations.

Acknowledgments

This model was developed as part of the MAP (Misconception Annotation Project) competition on Kaggle. Special thanks to the competition hosts and the Kaggle community for their support and collaboration.

How to Use This Model

To use this model for predicting math misconceptions:

1. Install the required libraries:

pip install transformers torch

2. Load the model and tokenizer:

from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_name = "Qwen3-8B-Math-Misconception-Classifier"
model = AutoModelForSequenceClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

3. Prepare your input data in the required format.

4. Make predictions:

from transformers import pipeline

classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)
results = classifier(your_input_data)

5. Interprete the results, which will include the predicted misconception categories and their associated probabilities.

References

• Hugging Face Transformers Documentation
• Kaggle MAP Competition