MASRISET-V4-DOWNSTREAM-SPLIT

Dataset Description

MASRISET-V4-DOWNSTREAM-SPLIT is a multi-task dataset for Egyptian Arabic natural language understanding. It combines three complementary tasks:

Task	Classes	Label Schema
Sentiment	3	0=Negative, 1=Neutral, 2=Positive
Emotion	4	0=Anger, 1=Joy, 2=Neutral, 3=Sadness
Sarcasm	2	0=Literal, 1=Sarcastic

This dataset is specifically designed for multi-task fine-tuning of language models, enabling simultaneous training on all three tasks with zero cross-task leakage.

Data Sources

Source	Task	Size	Notes
Egyptian Fake Reviews	Sentiment + Toxicity	~50K	E-commerce reviews with sentiment labels
Egyptian Sentiment Analysis	Sentiment	~40K	Social media sentiment
ASTD / ArSAS	Sentiment	~15K	General Arabic sentiment
ArSarcasm / iSarcasm	Sarcasm	~23K	Sarcasm detection in Arabic tweets
Emotone-AR	Emotion	~17K	8-class emotion mapped to 4-class schema
Arabic Hate Speech / Offenseval	Toxicity	~12K	Injected as Negative Sentiment + Anger/Sadness

Label Mapping for Emotion

The original Emotone-AR 8-class schema was mapped to a 4-class schema:

Original	Mapped	Notes
Anger (0)	Anger (0)	Direct
Fear/Disgust (1)	Sadness (3)	Negative valence
Joy (2)	Joy (1)	Direct
Neutral (3)	Neutral (2)	Direct
Sadness (4)	Sadness (3)	Direct
Surprise (5)	Neutral (2)	Ambiguous
Love (6)	Joy (1)	Positive valence
Sympathy (7)	Neutral (2)	Mixed

Toxicity Injection

Hate speech, offensive language, and toxic comments were injected with:

• Sentiment: Label = 0 (Negative)
• Emotion: Randomly assigned to Anger (0) or Sadness (3)

This ensures the model learns to recognize negative/aggressive language without needing a separate toxicity head.

Dataset Construction

1. Loading: All datasets loaded from Hugging Face Hub and standardized
2. Cleaning: clean_text_v4() applied uniformly (removes Latin, diacritics, URLs, etc.)
3. Deduplication: Conflicting labels resolved via text deduplication
4. Augmentation:
   • Sarcasm: Over-sampled minority class (2x)
   • Sentiment: Downsampled for balance
   • Emotion: Augmented 4x using word operations
5. Leakage Prevention: Validation and test sets rigorously purged of train-set texts
6. Global Split: 70/15/15 train/validation/test split

Dataset Statistics

Split	Size	Contains Tasks
Train	84,795	Sentiment (100%), Sarcasm (22K), Emotion (39K)
Validation	18,171	All three
Test	18,171	All three
Total	121,137

Class Distributions

Sentiment (Train):

• 2=Positive (42,871)
• 0=Negative (24,326)
• 1=Neutral (20,733)

Sarcasm (Train):

• 0=Literal (19,791)
• 1=Sarcastic (4,010)

Emotion (Train): Balanced across 4 classes (~9.8K each)

Data Format

{
  "text": "الخدمة دى وحشة جدا مش هتعامل معاكم تانى",
  "sentiment": 0,
  "emotion": 0,
  "sarcasm": 0
}

Note: Some entries may have missing labels (NaN) for certain tasks – this is intended for multi-task learning.

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("T0KII/MASRISET-V4-DOWNSTREAM-SPLIT")

# Access splits
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

# Example with multi-task training
for example in train:
    text = example["text"]
    sentiment_label = example["sentiment"]
    emotion_label = example["emotion"]
    sarcasm_label = example["sarcasm"]

For Multi-Task Fine-Tuning

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch.nn as nn

class MultiTaskHead(nn.Module):
    def __init__(self, base_model, num_sentiment=3, num_emotion=4, num_sarcasm=2):
        super().__init__()
        self.base = base_model
        hidden_size = base_model.config.hidden_size
        self.sentiment_head = nn.Linear(hidden_size, num_sentiment)
        self.emotion_head = nn.Linear(hidden_size, num_emotion)
        self.sarcasm_head = nn.Linear(hidden_size, num_sarcasm)
        
    def forward(self, input_ids, attention_mask):
        outputs = self.base(input_ids, attention_mask, output_hidden_states=True)
        pooled = outputs.last_hidden_state[:, 0]
        sentiment_logits = self.sentiment_head(pooled)
        emotion_logits = self.emotion_head(pooled)
        sarcasm_logits = self.sarcasm_head(pooled)
        return sentiment_logits, emotion_logits, sarcasm_logits

License

CC-BY-NC-4.0

Citation

@misc{masriset-v4-downstream,
  author = {T0KII},
  title = {MASRISET-V4-DOWNSTREAM-SPLIT: A Multi-Task Egyptian Arabic Dataset for Sentiment, Emotion, and Sarcasm},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/T0KII/MASRISET-V4-DOWNSTREAM-SPLIT}
}

Acknowledgements

• IbrahimAmin for Egyptian fake reviews and hate speech datasets
• ehab215 for Egyptian sentiment dataset
• ArbML team for ASTD, ArSAS, TEAD, and iSarcasm
• Emotone-AR team for the emotion dataset
• NoraAlt for sarcasm dataset