# ParsBigBird: Persian Bert For Long-Range Sequences

The Bert and ParsBert algorithms can handle texts with token lengths of up to 512, however, many tasks such as summarizing and answering questions require longer texts. In our work, we have trained the BigBird model for the Persian language to process texts up to 4096 in the Farsi (Persian) language using sparse attention.

## Evaluation: 🌡️

We have evaluated the model on three tasks with different sequence lengths

Name Params SnappFood (F1) Digikala Magazine(F1) PersianQA (F1)
distil-bigbird-fa-zwnj 78M 85.43% 94.05% 73.34%
bert-base-fa 118M 87.98% 93.65% 70.06%
• Despite being as big as distill-bert, the model performs equally well as ParsBert and is much better on PersianQA which requires much more context
• This evaluation was based on max_lentgh=2048 (It can be changed up to 4096)

## How to use❓

### As Contextualized Word Embedding

from transformers import BigBirdModel, AutoTokenizer

# by default its in block_sparse block_size=32
model = BigBirdModel.from_pretrained(MODEL_NAME, block_size=32)
# you can use full attention like the following: use this when input isn't longer than 512
model = BigBirdModel.from_pretrained(MODEL_NAME, attention_type="original_full")

text = "😃 امیدوارم مدل بدردبخوری باشه چون خیلی طول کشید تا ترین بشه"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
tokens = tokenizer(text, return_tensors='pt')
output = model(**tokens) # contextualized embedding


### As Fill Blank

from transformers import pipeline

results = fill('تهران پایتخت [MASK] است.')
print(results[0]['token_str'])
>>> 'ایران'


## Pretraining details: 🔭

This model was pretrained using a masked language model (MLM) objective on the Persian section of the Oscar dataset. Following the original BERT training, 15% of tokens were masked. This was first described in this paper and released in this repository. Documents longer than 4096 were split into multiple documents, while documents much smaller than 4096 were merged using the [SEP] token. Model is warm started from distilbert-fa’s checkpoint.

• For more details, you can take a look at config.json at the model card in 🤗 Model Hub

## Fine Tuning Recommendations: 🐤

Due to the model's memory requirements, gradient_checkpointing and gradient_accumulation should be used to maintain a reasonable batch size. Considering this model isn't really big, it's a good idea to first fine-tune it on your dataset using Masked LM objective (also called intermediate fine-tuning) before implementing the main task. In block_sparse mode, it doesn't matter how many tokens are input. It just attends to 256 tokens. Furthermore, original_full should be used up to 512 sequence lengths (instead of block sparse).

### Fine Tuning Examples 👷‍♂️👷‍♀️

Dataset Fine Tuning Example
Digikala Magazine Text Classification

If you have a technical question regarding the model, pretraining, code or publication, please create an issue in the repository. This is the fastest way to reach us.

## Citation: ↩️

we didn't publish any papers on the work. However, if you did, please cite us properly with an entry like one below.

@misc{ParsBigBird,
title           = {ParsBigBird: Persian Bert For Long-Range Sequences},
year            = 2021,
publisher       = {GitHub},
journal         = {GitHub repository},

Mask token: undefined