A fine-tuned XLM-RoBERTa model for 7-class emotion classification in Roman Urdu text, covering Ekman's six basic emotions plus a none class.

Model Details

Field	Value
Base model	`Khubaib01/roman-urdu-sentiment-xlm-r`
Task	Multi-class emotion classification
Language	Roman Urdu (Romanized Urdu / code-switched)
Classes	7 (anger, disgust, fear, joy, sadness, surprise, none)
Parameters	~278M (XLM-R base + custom head)
Max input length	128 tokens

Emotion Classes

Label	Ekman Emotion
anger	Basic
disgust	Basic
fear	Basic
joy	Basic
sadness	Basic
surprise	Basic
none	No emotion

Performance

Metrics calculated on test(unseen) dataset.

Metric	Score
Macro F1	0.7149
Accuracy	0.7148

Per-class F1 (Test Set)

Emotion	F1 Score
Anger	0.5962
Disgust	0.5672
Fear	0.8467
Happy	0.7276
Sad	0.6379
Surprise	0.9349
None	0.6937

Architecture

Input (Roman Urdu text)
    → XLM-R Tokenizer (max_length=128)
    → XLM-R Encoder (12 layers, hidden=768)
    → [CLS] token representation
    → LayerNorm → Dropout(0.3)
    → Linear(768 → 256) → GELU → Dropout(0.15)
    → Linear(256 → 7)
    → CrossEntropyLoss (label_smoothing=0.1)

Training Details

Hyperparameter	Value
Base LR (encoder)	2e-5
Head LR	1e-4
Layer-wise LR decay	0.95 per layer
Scheduler	Cosine with 10% warm-up
Epochs	Up to 10 (early stopping)
Batch size	16
Label smoothing	0.1
Dropout	0.3
Optimizer	AdamW
Mixed precision	fp16

Key technique: Discriminative layer-wise learning rates — the classification head trains at 5× the base encoder rate, while lower encoder layers decay further to preserve pre-trained multilingual representations.

Dataset

Total samples: 21,000
Split: 80% train / 10% val / 10% test
Balance: Perfectly balanced (3,000 samples per class)
Language: Roman Urdu (social media messages)

Usage

⚠️ Important — Custom Architecture

This model uses a custom PyTorch nn.Module, not a standard HuggingFace PreTrainedModel. You must define the class before loading weights.

python

from transformers import pipeline

HF_REPO_ID = "Khubaib01/roman-urdu-emotion-xlmr"

pipe = pipeline(
        "text-classification",
        model="f{HF_REPO_ID}",
        trust_remote_code=True,
        top_k=None,
    )
pipe("bhai ab mera kia hoga")

Limitations

Trained on a specific Roman Urdu dataset; performance may vary on highly code-switched (Urdu + English) or dialectal text.
The none class captures emotionally neutral text but may misfire on sarcasm or implicit sentiment.
Not evaluated on formal/literary Urdu.

Citation

If you use this model in your research, please cite:

bibtex

@misc{roman-urdu-emotion-xlmr,
  author    = {Muhammad Khubaib Ahmad},
  title     = {Roman Urdu Emotion Classifier (XLM-R)},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/Khubaib01/roman-urdu-emotion-xlmr}
}