FaBERT: Pre-training BERT on Persian Blogs
Model Details
FaBERT is a Persian BERT-base model trained on the diverse HmBlogs corpus, encompassing both casual and formal Persian texts. Developed for natural language processing tasks, FaBERT is a robust solution for processing Persian text. Through evaluation across various Natural Language Understanding (NLU) tasks, FaBERT consistently demonstrates notable improvements, while having a compact model size. Now available on Hugging Face, integrating FaBERT into your projects is hassle-free. Experience enhanced performance without added complexity as FaBERT tackles a variety of NLP tasks.
Features
- Pre-trained on the diverse HmBlogs corpus consisting more than 50 GB of text from Persian Blogs
- Remarkable performance across various downstream NLP tasks
- BERT architecture with 124 million parameters
Useful Links
- Repository: FaBERT on Github
- Paper: ACL Anthology
Usage
Loading the Model with MLM head
from transformers import AutoTokenizer, AutoModelForMaskedLM
tokenizer = AutoTokenizer.from_pretrained("sbunlp/fabert") # make sure to use the default fast tokenizer
model = AutoModelForMaskedLM.from_pretrained("sbunlp/fabert")
Downstream Tasks
Similar to the original English BERT, FaBERT can be fine-tuned on many downstream tasks.(https://huggingface.co/docs/transformers/en/training)
Examples on Persian datasets are available in our GitHub repository.
make sure to use the default Fast Tokenizer
Training Details
FaBERT was pre-trained with the MLM (WWM) objective, and the resulting perplexity on validation set was 7.76.
| Hyperparameter | Value |
|---|---|
| Batch Size | 32 |
| Optimizer | Adam |
| Learning Rate | 6e-5 |
| Weight Decay | 0.01 |
| Total Steps | 18 Million |
| Warmup Steps | 1.8 Million |
| Precision Format | TF32 |
Evaluation
Here are some key performance results for the FaBERT model:
Sentiment Analysis
| Task | FaBERT | ParsBERT | XLM-R |
|---|---|---|---|
| MirasOpinion | 87.51 | 86.73 | 84.92 |
| MirasIrony | 74.82 | 71.08 | 75.51 |
| DeepSentiPers | 79.85 | 74.94 | 79.00 |
Named Entity Recognition
| Task | FaBERT | ParsBERT | XLM-R |
|---|---|---|---|
| PEYMA | 91.39 | 91.24 | 90.91 |
| ParsTwiner | 82.22 | 81.13 | 79.50 |
| MultiCoNER v2 | 57.92 | 58.09 | 51.47 |
Question Answering
| Task | FaBERT | ParsBERT | XLM-R |
|---|---|---|---|
| ParsiNLU | 55.87 | 44.89 | 42.55 |
| PQuAD | 87.34 | 86.89 | 87.60 |
| PCoQA | 53.51 | 50.96 | 51.12 |
Natural Language Inference & QQP
| Task | FaBERT | ParsBERT | XLM-R |
|---|---|---|---|
| FarsTail | 84.45 | 82.52 | 83.50 |
| SBU-NLI | 66.65 | 58.41 | 58.85 |
| ParsiNLU QQP | 82.62 | 77.60 | 79.74 |
Number of Parameters
| FaBERT | ParsBERT | XLM-R | |
|---|---|---|---|
| Parameter Count (M) | 124 | 162 | 278 |
| Vocabulary Size (K) | 50 | 100 | 250 |
For a more detailed performance analysis refer to the paper.
How to Cite
If you use FaBERT in your research or projects, please cite it using the following BibTeX:
@inproceedings{masumi-etal-2025-fabert,
title = "{F}a{BERT}: Pre-training {BERT} on {P}ersian Blogs",
author = "Masumi, Mostafa and
Majd, Seyed Soroush and
Shamsfard, Mehrnoush and
Beigy, Hamid",
editor = "Bak, JinYeong and
Goot, Rob van der and
Jang, Hyeju and
Buaphet, Weerayut and
Ramponi, Alan and
Xu, Wei and
Ritter, Alan",
booktitle = "Proceedings of the Tenth Workshop on Noisy and User-generated Text",
month = may,
year = "2025",
address = "Albuquerque, New Mexico, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.wnut-1.10/",
doi = "10.18653/v1/2025.wnut-1.10",
pages = "85--96",
ISBN = "979-8-89176-232-9",
}