|
--- |
|
language: |
|
- he |
|
tags: |
|
- language model |
|
license: apache-2.0 |
|
datasets: |
|
- oscar |
|
- wikipedia |
|
- twitter |
|
--- |
|
|
|
# AlephBERT |
|
|
|
## Hebrew Language Model |
|
|
|
State-of-the-art language model for Hebrew. |
|
Based on Google's BERT architecture [(Devlin et al. 2018)](https://arxiv.org/abs/1810.04805). |
|
|
|
#### How to use |
|
|
|
```python |
|
from transformers import BertModel, BertTokenizerFast |
|
|
|
alephbert_tokenizer = BertTokenizerFast.from_pretrained('onlplab/alephbert-base') |
|
alephbert = BertModel.from_pretrained('onlplab/alephbert-base') |
|
|
|
# if not finetuning - disable dropout |
|
alephbert.eval() |
|
``` |
|
|
|
## Training data |
|
1. OSCAR [(Ortiz, 2019)](https://oscar-corpus.com/) Hebrew section (10 GB text, 20 million sentences). |
|
2. Hebrew dump of [Wikipedia](https://dumps.wikimedia.org/hewiki/latest/) (650 MB text, 3 million sentences). |
|
3. Hebrew Tweets collected from the Twitter sample stream (7 GB text, 70 million sentences). |
|
|
|
## Training procedure |
|
|
|
Trained on a DGX machine (8 V100 GPUs) using the standard huggingface training procedure. |
|
|
|
Since the larger part of our training data is based on tweets we decided to start by optimizing using Masked Language Model loss only. |
|
|
|
To optimize training time we split the data into 4 sections based on max number of tokens: |
|
|
|
1. num tokens < 32 (70M sentences) |
|
2. 32 <= num tokens < 64 (12M sentences) |
|
3. 64 <= num tokens < 128 (10M sentences) |
|
4. 128 <= num tokens < 512 (1.5M sentences) |
|
|
|
Each section was first trained for 5 epochs with an initial learning rate set to 1e-4. Then each section was trained for another 5 epochs with an initial learning rate set to 1e-5, for a total of 10 epochs. |
|
|
|
Total training time was 8 days. |
|
|
|
|
|
|
