abrhamLearns/afriaware-amharic-hate-speech

AfriAware — Amharic Hate Speech Classifier 🛡️

A fine-tuned Amharic hate speech classifier for digital safety research in Ethiopia and across Africa.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="YOUR_USERNAME/afriaware-amharic-hate-speech"
)

result = classifier("ኢትዮጵያ ሀገራችን ናት")
# [{'label': 'normal', 'score': 0.94}]

Labels

Label	ID	Meaning
hate	0	Hatred directed at a person or group
abusive	1	Offensive/aggressive language
normal	2	Non-harmful content

Training Details

Setting	Value
Base model	Davlan/bert-base-multilingual-cased-finetuned-amharic
Dataset	AfriHate (amh)
Epochs	5 (early stopping, patience=2)
Learning rate	2e-5
Batch size	16
Max length	128 tokens
Optimizer	AdamW, weight decay 0.01

Dataset

Trained on the Amharic subset of AfriHate (Muhammad et al., NAACL 2025) — tweets annotated by native Amharic speakers with full sociocultural context.

Ethical Considerations

• Intended for research and journalist support tools, not automated removal
• Annotations reflect Ethiopian social media from 2020–2022
• Model should always be combined with human review
• Do not use for surveillance of individuals

Citation

@inproceedings{muhammad-etal-2025-afrihate,
    title = "{A}fri{H}ate: A Multilingual Collection of Hate Speech and Abusive
             Language Datasets for {A}frican Languages",
    author = {Muhammad, Shamsuddeen Hassan and others},
    booktitle = "Proceedings of NAACL 2025",
    year = "2025"
}