Eduagent_distilbert

DistilBERT EduAgent v2 — Fine-Tuned Difficulty Classifier

This directory contains the fine-tuned DistilBERT model for predicting learner difficulty levels (beginner, intermediate, advanced) in AI/ML educational contexts.

Model Overview

• Base model: distilbert-base-uncased (6 layers, 66M parameters)
• Training dataset: 2,400 balanced synthetic AI/ML Q&A pairs (800 per difficulty level)
• Train/val/test split: 1,920 / 240 / 240 samples
• Test accuracy: 97.92% (+35.8 pp over keyword heuristic baseline of 62.1%)
• Inference: CPU-only, < 100 ms per prediction
• Fine-tuning framework: HuggingFace Transformers

Files

config.json                    ← Model config (vocab size, hidden dims, etc.)
model.safetensors             ← Model weights (safetensors format)
tokenizer_config.json         ← Tokenizer config
special_tokens_map.json       ← Special token mappings
vocab.txt                     ← Vocabulary

Usage

In EduAgent

The model is loaded automatically by ml/classifier.py:

from ml.classifier import classify_difficulty

level, confidence = classify_difficulty("What is supervised learning?")
# Returns: ("beginner", [0.98, 0.01, 0.01])

Standalone (HuggingFace Hub)

from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "Sneha-260805/distilbert-eduagent-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

inputs = tokenizer("What is a neural network?", return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
predicted_level = logits.argmax(dim=-1).item()
# 0 = beginner, 1 = intermediate, 2 = advanced

Performance

Test Set Results

Confusion Matrix (Test Set)

        Predicted Beginner  Intermediate  Advanced
Actual  
Beginner      79              1             0
Inter.         0              78             2
Advanced       0              3             77

Comparison with Baseline

Training Details

• Framework: HuggingFace Trainer
• Loss: CrossEntropyLoss
• Optimizer: AdamW with weight decay
• Learning rate: 2e-5
• Batch size: 16
• Epochs: 3
• Warmup steps: 100
• Device: CUDA (GPU) during training; CPU/GPU at inference

Fallback Behavior

If this model directory is absent, ml/classifier.py automatically falls back to a lightweight keyword heuristic classifier:

def fallback_classifier(text):
    """Keyword-based difficulty detection (62.1% accuracy)"""
    text_lower = text.lower()
    beginner_keywords = ["what is", "explain", "basics", "introduction", "simple"]
    intermediate_keywords = ["how to", "implement", "optimize", "design pattern"]
    advanced_keywords = ["edge case", "architectural", "performance tuning", "novel"]
    # ... scoring logic ...

This ensures the app remains functional even if model weights are unavailable, though with significantly lower accuracy.

Integration with EduAgent

The classifier is called at the start of every tutor interaction:

1. User asks a question
2. classify_difficulty(question) returns (level, [logits])
3. Confidence scores drive memory hints (low confidence → more conservative teaching)
4. Level is recorded in learner history for trend analysis
5. Dataset retrieval is filtered to the detected level

Model Limitations

• Training domain: Synthetic AI/ML Q&A pairs — may not generalize perfectly to other educational domains (e.g., biology, history)
• Input length: Optimized for questions ≤ 512 tokens (DistilBERT max)
• Balance: Fine-tuned on balanced dataset (800 per class) — may have bias on extremely rare difficulty distributions
• Real-time feedback: Trained on static question snapshots — does not adapt to individual learner history

Citation

If you use this model in your research or project, please cite:

@misc{eduagent2024,
  author = {Suravajjula, Sneha},
  title = {EduAgent: Adaptive AI Tutor with Personalized Learning Loop},
  year = {2024},
  url = {https://github.com/Sneha-260805/EduAgent}
}

License

This model is released under the same license as the EduAgent project. See the root repository LICENSE for details.

HuggingFace Hub

Model weights are also published at:

Sneha-260805/distilbert-eduagent-v2

This allows direct loading via HuggingFace transformers without cloning the entire EduAgent repository.

---

For full system architecture, training details, and experimental results, see report.tex.