EthicsBERT

EthicsBERT is a fine-tuned DistilBERT model for classifying AI ethics course content into nine topic areas. It is designed to support educators, researchers, and curriculum designers working in AI ethics.

Labels

ID	Label	Description
0	Agency	Human control, override rights, autonomy-preserving design, and delegated authority.
1	AI Governance	Regulation, accountability, audits, policy frameworks, and existential risk oversight.
2	Bias	Systematic errors, skewed training data, unfair representations, and proxy discrimination.
3	Consciousness	Machine sentience, subjective experience, philosophical debates, and moral patienthood.
4	Ethical Reasoning	Moral frameworks (utilitarian, deontological, virtue), dilemmas, and applied ethics.
5	Explainability	Interpretability, SHAP/LIME, attention visualization, and model transparency.
6	Fairness	Equitable outcomes, anti-discrimination, group/individual fairness metrics.
7	Intelligence	Cognitive capabilities, reasoning, transfer learning, AGI, and benchmarks.
8	Privacy	Data protection, consent, PII handling, differential privacy, and encryption.

Model Details

Property	Value
Base model	distilbert-base-uncased
Architecture	DistilBERT + classification head
Task	Multi-class text classification (9 classes)
Max sequence length	128 tokens
Training epochs	5 (with early stopping)
Optimizer	AdamW
Learning rate	2e-5
Weight decay	0.01
Warmup ratio	0.1

Usage

With the pipeline API (simplest)

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="nexageapps/EthicsBERT",
    top_k=3,
)

result = classifier("The hiring algorithm must produce equal outcomes across demographic groups.")
# [{'label': 'Fairness', 'score': 0.92}, ...]
print(result)

Direct usage

import torch
import torch.nn.functional as F
from transformers import DistilBertForSequenceClassification, DistilBertTokenizerFast

model_id = "nexageapps/EthicsBERT"
tokenizer = DistilBertTokenizerFast.from_pretrained(model_id)
model = DistilBertForSequenceClassification.from_pretrained(model_id)
model.eval()

text = "SHAP values quantify each feature's contribution to a specific model prediction."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    probs = F.softmax(logits, dim=-1)

id2label = model.config.id2label
predicted_label = id2label[int(probs.argmax())]
print(f"Predicted: {predicted_label} ({probs.max().item():.2%})")

Training Data

The model was fine-tuned on a curated dataset of approximately 200 sentences covering all nine AI ethics topic areas. Sentences were authored to reflect the vocabulary, framing, and conceptual depth typical of AI ethics course materials.

Label distribution: roughly balanced, ~20–25 examples per class.

Limitations

• The training dataset is small (~200 examples). Performance may degrade on highly technical or domain-specific text not represented in training.
• The model was trained on English only.
• Boundary cases between semantically similar labels (e.g., Fairness vs Bias, or AI Governance vs Ethical Reasoning) may be misclassified.
• The model should not be used as the sole arbiter in automated grading or gatekeeping systems.

Ethical Considerations

This model is intended for research and educational purposes. Automated topic classification of ethics content should always be reviewed by a human expert before consequential use.

Citation

If you use EthicsBERT in your research or course materials, please cite:

@misc{ethicsbert2024,
  title   = {EthicsBERT: A DistilBERT Model for AI Ethics Topic Classification},
  author  = {nexageapps},
  year    = {2024},
  url     = {https://huggingface.co/nexageapps/EthicsBERT}
}

License

Apache 2.0 — see LICENSE for details.