EthicsBERT

A fine-tuned DistilBERT model for classifying AI ethics course content into nine topic areas. Built to support educators, researchers, and curriculum designers.

Labels

Label	Description
Agency	Human control, override rights, and autonomy-preserving AI design
AI Governance	Regulation, audits, policy, accountability, and existential risk
Bias	Systematic errors, skewed data, proxy discrimination
Consciousness	Machine sentience, subjective experience, philosophical debates
Ethical Reasoning	Moral frameworks, dilemmas, applied ethics principles
Explainability	SHAP/LIME, interpretability, attention, model transparency
Fairness	Equitable outcomes, anti-discrimination, parity metrics
Intelligence	Reasoning, transfer learning, AGI, cognitive benchmarks
Privacy	Data protection, consent, PII, differential privacy

Project Structure

EthicsClassifierModel/
├── EthicsBERT/
│   ├── data/
│   │   └── sample_ethics_dataset.csv   # ~200 labelled training examples
│   ├── scripts/
│   │   ├── train.py                    # Fine-tune DistilBERT
│   │   ├── evaluate.py                 # Full classification report
│   │   ├── infer.py                    # Single / batch inference
│   │   └── deploy.py                   # Push to Hugging Face Hub
│   ├── app.py                          # Gradio demo (HF Spaces ready)
│   └── MODEL_CARD.md                   # Hugging Face model card
├── requirements.txt
└── README.md

Setup

python3 -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

# Core ML deps (train, evaluate, infer, deploy)
pip install -r requirements.txt

# Gradio demo only (separate due to huggingface_hub version split)
pip install -r requirements-app.txt

Train

python3 EthicsBERT/scripts/train.py \
  --dataset_path EthicsBERT/data/sample_ethics_dataset.csv \
  --output_dir   EthicsBERT/model \
  --epochs       5 \
  --batch_size   16 \
  --learning_rate 2e-5

Key flags:

Flag	Default	Description
--dataset_path	EthicsBERT/data/sample_ethics_dataset.csv	Training CSV
--output_dir	EthicsBERT/model	Where to save the model
--model_name	distilbert-base-uncased	Base HF checkpoint
--epochs	5	Number of training epochs
--batch_size	16	Per-device batch size
--learning_rate	2e-5	AdamW learning rate
--early_stopping_patience	2	Stop if no F1 improvement

Evaluate

python3 EthicsBERT/scripts/evaluate.py \
  --dataset_path EthicsBERT/data/sample_ethics_dataset.csv \
  --model_dir    EthicsBERT/model \
  --output_report EthicsBERT/eval_report.txt

Inference

Single text:

python3 EthicsBERT/scripts/infer.py \
  --model_dir EthicsBERT/model \
  --text "The board requested clear model documentation and audit trails."

Batch from file (one sentence per line):

python3 EthicsBERT/scripts/infer.py \
  --model_dir   EthicsBERT/model \
  --input_file  my_texts.txt \
  --top_k       3

Using a Hub-hosted model directly:

python3 EthicsBERT/scripts/infer.py \
  --model_dir nexageapps/EthicsBERT \
  --text "Differential privacy protects individuals in aggregate queries."

Deploy to Hugging Face Hub

# Authenticate once
huggingface-cli login

# Push model + tokenizer + model card
python3 EthicsBERT/scripts/deploy.py \
  --model_dir       EthicsBERT/model \
  --repo_id         nexageapps/EthicsBERT \
  --model_card_path EthicsBERT/MODEL_CARD.md

Gradio Demo (Local or HF Spaces)

pip install gradio
python3 EthicsBERT/app.py

To deploy to Hugging Face Spaces:

1. Create a new Space (Gradio SDK).
2. Push EthicsBERT/app.py and requirements.txt to the Space repo.
3. Set HF_MODEL_ID=nexageapps/EthicsBERT in the Space secrets.

Quick API Usage (after Hub deployment)

from transformers import pipeline

clf = pipeline("text-classification", model="nexageapps/EthicsBERT", top_k=3)
print(clf("SHAP values explain each feature's contribution to the prediction."))

Requirements

• Python ≥ 3.10
• PyTorch ≥ 2.0
• Transformers ≥ 4.41
• See requirements.txt for full pinned dependency list