Khayyam — Persian Multi-Label Emotion Classifier
Khayyam is a fine-tuned Persian BERT model for fine-grained, multi-label emotion classification across 28 emotion categories, based on the GoEmotions taxonomy. It is built on top of ParsBERT v3 and trained on GoEmotions Persian — a Persian translation of Google's GoEmotions benchmark dataset.
The model is named after Omar Khayyam, the 11th-century Persian poet, mathematician, and philosopher.
Model Details
| Property | Value |
|---|---|
| Base model | HooshvareLab/bert-fa-zwnj-base (ParsBERT v3) |
| Task | Multi-label text classification |
| Language | Persian (Farsi / fa) |
| Number of labels | 28 |
| Decision threshold | 0.3 |
| License | Apache 2.0 |
Evaluation Results
Evaluated on the held-out test split of GoEmotions Persian (5,427 samples):
| Metric | Score |
|---|---|
| F1 Macro | 0.4124 |
| F1 Micro | 0.5563 |
For reference, the original GoEmotions paper (Demszky et al., 2020) reports F1 Macro of 0.46 for BERT-base on the English dataset. Khayyam achieves competitive results considering the additional challenges of translation noise and Persian-specific linguistic features.
Training curve
| Epoch | Train Loss | Val Loss | F1 Macro | F1 Micro |
|---|---|---|---|---|
| 1 | 0.1029 | 0.1026 | 0.3231 | 0.5313 |
| 2 | 0.0934 | 0.0950 | 0.3817 | 0.5597 |
| 3 | 0.0826 | 0.0939 | 0.4016 | 0.5585 |
| 4 | 0.0768 | 0.0951 | 0.4105 | 0.5574 |
Usage
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model_name = "aydakikio/Khayyam"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
LABELS = [
"admiration", "amusement", "anger", "annoyance", "approval", "caring",
"confusion", "curiosity", "desire", "disappointment", "disapproval",
"disgust", "embarrassment", "excitement", "fear", "gratitude", "grief",
"joy", "love", "nervousness", "optimism", "pride", "realization",
"relief", "remorse", "sadness", "surprise", "neutral"
]
THRESHOLD = 0.3
def predict_emotions(text):
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
logits = model(**inputs).logits
probs = torch.sigmoid(logits).squeeze()
predicted = [LABELS[i] for i, p in enumerate(probs) if p >= THRESHOLD]
return predicted
# Example
text = "امروز خیلی خوشحالم، همه چیز عالی پیش رفت!"
print(predict_emotions(text))
# → ['joy', 'optimism', 'excitement']
Training Details
Dataset
- Dataset: GoEmotions Persian
- Source: Persian translation of Google GoEmotions using the Gap GPT API
- Size: 54,263 samples (train: 43,410 / dev: 5,426 / test: 5,427)
- Labels: 27 emotion categories + Neutral (28 total)
- License: CC BY 4.0
Hyperparameters
| Parameter | Value |
|---|---|
| Epochs | 4 |
| Learning rate | 2e-5 |
| Batch size | 16 |
| Max sequence length | 128 |
| Warmup ratio | 0.1 |
| Weight decay | 0.01 |
| Optimizer | AdamW |
| Loss function | Binary Cross-Entropy (multi-label) |
| Precision | fp16 |
Emotion Labels
The model classifies text into any combination of these 28 emotions:
admiration · amusement · anger · annoyance · approval · caring · confusion · curiosity · desire · disappointment · disapproval · disgust · embarrassment · excitement · fear · gratitude · grief · joy · love · nervousness · optimism · pride · realization · relief · remorse · sadness · surprise · neutral
Limitations
- Trained on translated data — some label noise may exist due to automatic translation
- Performance on rare emotion classes (e.g.,
grief,pride,relief) is lower than on frequent ones - Optimized for Persian Reddit-style short text; may generalize differently to formal or dialectal Persian
Citation
If you use this model, please cite:
Version 1.0
@misc{ayda_khoshkhan_2026,
author = { Ayda Khoshkhan },
title = { Khayyam (Revision 4cbc751) },
year = 2026,
url = { https://huggingface.co/aydakikio/Khayyam },
doi = { 10.57967/hf/9089 },
publisher = { Hugging Face }
}
Also cite the base model and dataset:
@article{ParsBERT,
title = {ParsBERT: Transformer-based Model for Persian Language Understanding},
author = {Mehrdad Farahani and Mohammad Gharachorloo and Marzieh Farahani and Mohammad Manthouri},
journal = {Neural Processing Letters},
year = {2021},
publisher = {Springer}
}
@inproceedings{demszky-etal-2020-goemotions,
title = {{G}o{E}motions: A Dataset of Fine-Grained Emotions},
author = {Demszky, Dorottya and Movshovitz-Attias, Dana and Ko, Jeongwoo and Cowen, Alan and Nemade, Gaurav and Ravi, Sujith},
booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
year = {2020},
publisher = {Association for Computational Linguistics}
}
License
This model is released under the Apache 2.0 license, consistent with the base model (HooshvareLab/bert-fa-zwnj-base) and the original GoEmotions dataset.
The training dataset (aydakikio/goemotion_persian) is licensed under CC BY 4.0.