# BERT-based Multi-label Cognitive Load Classifier

This model is a fine-tuned `bert-base-uncased` transformer trained to classify **students' cognitive and psychological states** (e.g., cognitive load, confidence, anxiety) from naturalistic **human-AI educational dialogues** in K-12 settings.

## 🧠 What does the model do?

The model performs **multi-label classification** on student-AI conversations, identifying whether a given interaction reflects one or more of the following cognitive and affective states:

- Math Confidence / Math Anxiety
- AI Confidence / AI Concerns
- Intrinsic Cognitive Load
- Extraneous Cognitive Load
- Germane Cognitive Load

Each input text (a single conversation) may correspond to **multiple labels simultaneously**.

---

## 📚 Training Data

The model was trained on a custom dataset collected from a large-scale empirical study involving **160 K-12 students** interacting with an AI-powered teachable agent in a math learning platform (ALTER-Math, name anonymized for review).

- **Dialogues**: 1,440 student-agent interactions over 10 days
- **Labels**: Derived from pre- and post-questionnaires grounded in Cognitive Load Theory and affective constructs
- **Label types**: Binary indicators (0/1) per psychological factor
- **Preprocessing**: Tokenized using Hugging Face's `AutoTokenizer`, padded to max length of 128

---

## 🏋️‍♂️ Training Setup

- Model: `bert-base-uncased`
- Task: Multi-label text classification
- Loss: BCEWithLogitsLoss
- Optimizer: AdamW
- Batch Size: 16
- Epochs: 5
- Learning Rate: 1e-5
- Evaluation Strategy: Hold-out test set (20%)

---

## 🚀 Intended Use

This model is designed to support **AI-based unobtrusive assessment of cognitive load** in education, enabling:

- Researchers to monitor how students respond cognitively and emotionally to AI tutors
- Developers to build more adaptive, trustworthy AI learning agents
- Teachers to gain insight into student engagement and overload without invasive devices

---

## 📌 Limitations

- The dataset size is modest (N=160), and model generalization to other domains or age groups is not guaranteed.
- Labels are inferred from questionnaire-aligned criteria, which may include subjectivity.
- The model does not currently handle out-of-distribution input or code-switching effectively.