Model Details
TookaBERT models are a family of encoder models trained on Persian in two sizes base and large. These Models pre-trained on over 500GB of Persian data including a variety of topics such as News, Blogs, Forums, Books, etc. They pre-trained with the MLM (WWM) objective using two context lengths. TookaBERT-Large is the first large encoder model pre-trained on Persian and currently is the state-of-the-art model in Persian tasks.
For more information you can read our paper on arXiv.
How to use
You can use this model directly for Masked Language Modeling using the provided code below.
from transformers import AutoTokenizer, AutoModelForMaskedLM
tokenizer = AutoTokenizer.from_pretrained("PartAI/TookaBERT-Large")
model = AutoModelForMaskedLM.from_pretrained("PartAI/TookaBERT-Large")
# prepare input
text = "شهر برلین در کشور <mask> واقع شده است."
encoded_input = tokenizer(text, return_tensors='pt')
# forward pass
output = model(**encoded_input)
It is also possible to use inference pipelines such as below.
from transformers import pipeline
inference_pipeline = pipeline('fill-mask', model="PartAI/TookaBERT-Large")
inference_pipeline("شهر برلین در کشور <mask> واقع شده است.")
You can use this model to fine-tune it over your dataset and prepare it for your task.
- DeepSentiPers (Sentiment Analysis)
- ParsiNLU - Multiple-choice (Multiple-choice)
Evaluation
TookaBERT models are evaluated on a wide range of NLP downstream tasks, such as Sentiment Analysis (SA), Text Classification, Multiple-choice, Question Answering, and Named Entity Recognition (NER). Here are some key performance results:
| Model name | DeepSentiPers (f1/acc) | MultiCoNER-v2 (f1/acc) | PQuAD (best_exact/best_f1/HasAns_exact/HasAns_f1) | FarsTail (f1/acc) | ParsiNLU-Multiple-choice (f1/acc) | ParsiNLU-Reading-comprehension (exact/f1) | ParsiNLU-QQP (f1/acc) |
|---|---|---|---|---|---|---|---|
| TookaBERT-large | 85.66/85.78 | 69.69/94.07 | 75.56/88.06/70.24/87.83 | 89.71/89.72 | 36.13/35.97 | 33.6/60.5 | 82.72/82.63 |
| TookaBERT-base | 83.93/83.93 | 66.23/93.3 | 73.18/85.71/68.29/85.94 | 83.26/83.41 | 33.6/33.81 | 20.8/42.52 | 81.33/81.29 |
| Shiraz | 81.17/81.08 | 59.1/92.83 | 65.96/81.25/59.63/81.31 | 77.76/77.75 | 34.73/34.53 | 17.6/39.61 | 79.68/79.51 |
| ParsBERT | 80.22/80.23 | 64.91/93.23 | 71.41/84.21/66.29/84.57 | 80.89/80.94 | 35.34/35.25 | 20/39.58 | 80.15/80.07 |
| XLM-V-base | 83.43/83.36 | 58.83/92.23 | 73.26/85.69/68.21/85.56 | 81.1/81.2 | 35.28/35.25 | 8/26.66 | 80.1/79.96 |
| XLM-RoBERTa-base | 83.99/84.07 | 60.38/92.49 | 73.72/86.24/68.16/85.8 | 82.0/81.98 | 32.4/32.37 | 20.0/40.43 | 79.14/78.95 |
| FaBERT | 82.68/82.65 | 63.89/93.01 | 72.57/85.39/67.16/85.31 | 83.69/83.67 | 32.47/32.37 | 27.2/48.42 | 82.34/82.29 |
| mBERT | 78.57/78.66 | 60.31/92.54 | 71.79/84.68/65.89/83.99 | 82.69/82.82 | 33.41/33.09 | 27.2/42.18 | 79.19/79.29 |
| AriaBERT | 80.51/80.51 | 60.98/92.45 | 68.09/81.23/62.12/80.94 | 74.47/74.43 | 30.75/30.94 | 14.4/35.48 | 79.09/78.84 |
*Note because of the randomness in the fine-tuning process, results with less than 1% differences are considered together.
Contact us
If you have any questions regarding this model, you can reach us via the community of the model in Hugging Face.