You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to this dataset is gated. It contains hateful, offensive and abusive language. You agree to use it only for research and for building systems that detect, moderate or reduce hate speech, and not to use it to harass, target or harm any person or group.
Log in or Sign Up to review the conditions and access this dataset content.
Saraiki Hate Speech Dataset
A five-class text classification dataset for detecting hate speech and offensive language in Saraiki (Jataki/Multani), provided in Shahmukhi script with Roman transliteration.
Content warning: This dataset contains hateful, abusive, sexually explicit and offensive language, including content targeting people by religion and gender. It is provided only for research on detection and moderation.
Labels
| ID | Label | Description |
|---|---|---|
| 0 | Abusive/Offensive |
Insults or abuse directed at a person or group, without a religious or gender focus |
| 1 | Normal |
Neutral or benign text, with no hate or offensive content |
| 2 | Religious Hate |
Hate, takfir or abuse targeting people based on religion or sect |
| 3 | Sexism |
Abuse targeting people based on gender or sexuality, including gendered slurs |
| 4 | Profane/Untargeted |
Profanity or vulgarity not aimed at any person or group |
Quick start
The dataset is gated. Accept the conditions on this page, then log in with a Hugging Face token (huggingface-cli login).
# pip install -U datasets
from datasets import load_dataset
ds = load_dataset("themohal/saraiki-hate-speech")
print(ds)
print(ds["train"][0]["label_name"])
Baseline: fine-tune a classifier
# pip install -U transformers datasets accelerate scikit-learn
import numpy as np
from datasets import load_dataset
from sklearn.metrics import accuracy_score, f1_score
from transformers import (AutoModelForSequenceClassification, AutoTokenizer,
Trainer, TrainingArguments)
MODEL = "xlm-roberta-base" # any multilingual encoder
ID2LABEL = {0: "Abusive/Offensive", 1: "Normal", 2: "Religious Hate",
3: "Sexism", 4: "Profane/Untargeted"}
ds = load_dataset("themohal/saraiki-hate-speech")
tok = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForSequenceClassification.from_pretrained(
MODEL, num_labels=5, id2label=ID2LABEL, label2id={v: k for k, v in ID2LABEL.items()})
def encode(batch):
return tok(batch["saraiki"], truncation=True, max_length=128)
ds = ds.map(encode, batched=True).rename_column("label", "labels")
def metrics(pred):
preds = np.argmax(pred.predictions, axis=-1)
return {"accuracy": accuracy_score(pred.label_ids, preds),
"macro_f1": f1_score(pred.label_ids, preds, average="macro")}
args = TrainingArguments("out", per_device_train_batch_size=16, learning_rate=2e-5,
num_train_epochs=3, eval_strategy="epoch", report_to="none")
trainer = Trainer(model=model, args=args, train_dataset=ds["train"],
eval_dataset=ds["validation"], processing_class=tok,
compute_metrics=metrics)
trainer.train()
print(trainer.evaluate(ds["test"]))
Evaluation and reporting
- Report macro-F1 alongside accuracy, since the classes are likely imbalanced.
- Report per-class precision and recall. The
Religious Hate,SexismandProfane/Untargetedclasses are easy to confuse withAbusive/Offensive. - Use the
testsplit only for final evaluation, not for tuning.
Intended use
- Training and evaluating hate speech and offensive language detection for Saraiki.
- Content moderation and trust & safety research.
- Studying hate speech and code-mixing in low-resource languages.
Out-of-scope use
- Generating hateful, abusive or sexually explicit content.
- Harassing, profiling or targeting individuals or communities.
- Fully automated moderation decisions without human review.
- Any commercial use (see license).
Limitations and biases
- The Saraiki text was produced from Roman Urdu source text. It may differ from naturally occurring Saraiki, and some Urdu phrasing may carry over.
- Labels are inherited from the source dataset and may contain noise. Some examples have labels that do not clearly match their content (for example, general profanity labeled as religious hate), and some tweets lose context without the surrounding thread.
- Hate speech is context-dependent and culturally specific, and annotation involves subjective judgment.
- Class boundaries overlap. Texts can contain several kinds of abuse, but each has a single label.
- Models trained on this data may over-flag certain dialects, names or religious terms, including benign religious discussion. Evaluate for bias before deployment.
Access requirements
Access is gated. You must describe your intended use and agree to the usage conditions before accessing the dataset. By requesting access, you agree to use it only for non-commercial research and moderation purposes, not to use it to harm any person or group, and to cite it in any work that uses it.
Citation
If you use this dataset in research, publications, models, or public projects, please cite it:
@misc{raza2026saraikihatespeech,
title = {Saraiki Hate Speech Dataset},
author = {Raza, Muhammad Farjad Ali},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/themohal/saraiki-hate-speech}}
}
Disclaimer
This dataset is provided for research and educational purposes. Users are responsible for ensuring that their use complies with applicable laws, ethical requirements, and the dataset's license and access conditions. The views expressed in the data do not represent the views of the authors.
- Downloads last month
- 5,519