math-misconception-llama-3

🎓 Math Misconception Classifier (Llama-3 8B)

This model is fine-tuned to identify specific mathematical misconceptions in student explanations. It was developed as part of a Final Year Computer Science Major Project.

🚀 Model Details

• Developed by: Priyanshu Bhusan
• Model Type: Large Language Model (Fine-tuned for Classification)
• Base Model: Meta Llama-3 8B (4-bit Quantized)
• Task: Mapping student math misunderstandings to specific categories.
• Language: English
• Fine-tuning Technique: LoRA (Low-Rank Adaptation) via Unsloth.

📊 Performance Metrics

Based on the validation set (MAP Kaggle dataset):

• Top-1 Accuracy: 88.00%
• MAP@3 Score: 0.9090
• Weighted F1-Score: 0.7448

🛠️ Training Setup

• Hardware: NVIDIA L4/T4 GPU (Lightning AI Studio)
• Optimization: Unsloth 4-bit kernels for memory efficiency.
• Training Steps: 60 steps (Initial Trial)
• Learning Rate: 2e-4 with Linear Scheduler.

🧠 How to Use

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "vijigishu/math-misconception-llama-3",
    max_seq_length = 512,
    load_in_4bit = True,
)