Hugging Face's logo

allegro
/
herbert-klej-cased-v1

Transformers
PyTorch
JAX
Polish
roberta
Inference Endpoints
Model card Files Community
herbert-klej-cased-v1
File size: 4,012 Bytes 
06e3959
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6953ff8
06e3959
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6953ff8
 
 
 
 
 
 
 
 
 
 
 
 
 
06e3959
 
 
 
6953ff8
06e3959
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
 
---
language: pl
---

# HerBERT 
**[HerBERT](https://en.wikipedia.org/wiki/Zbigniew_Herbert)** is a BERT-based Language Model trained on Polish Corpora
using only MLM objective with dynamic masking of whole words. For more details, please refer to: 
[KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://arxiv.org/abs/2005.00630).

## Dataset
**HerBERT** training dataset is a combination of several publicly available corpora for Polish language:

| Corpus | Tokens | Texts |
| :------ | ------: | ------: |
| [OSCAR](https://traces1.inria.fr/oscar/)| 6710M  | 145M |
| [Open Subtitles](http://opus.nlpl.eu/OpenSubtitles-v2018.php) | 1084M  | 1.1M |
| [Wikipedia](https://dumps.wikimedia.org/) | 260M  | 1.5M |
| [Wolne Lektury](https://wolnelektury.pl/) | 41M  | 5.5k |
| [Allegro Articles](https://allegro.pl/artykuly) | 18M  | 33k |

## Tokenizer
The training dataset was tokenized into subwords using [HerBERT Tokenizer](https://huggingface.co/allegro/herbert-klej-cased-tokenizer-v1); a character level byte-pair encoding with
a vocabulary size of 50k tokens. The tokenizer itself was trained on [Wolne Lektury](https://wolnelektury.pl/) and a publicly available subset of 
[National Corpus of Polish](http://nkjp.pl/index.php?page=14&lang=0) with a [fastBPE](https://github.com/glample/fastBPE) library.

Tokenizer utilizes `XLMTokenizer` implementation for that reason, one should load it as `allegro/herbert-klej-cased-tokenizer-v1`.

## HerBERT models summary
| Model | WWM | Cased | Tokenizer | Vocab Size  | Batch Size | Train Steps |
| :------ | ------: | ------: | ------: | ------: | ------: | ------: |
| herbert-klej-cased-v1 | YES | YES | BPE | 50K | 570 | 180k | 

## Model evaluation
HerBERT was evaluated on the [KLEJ](https://klejbenchmark.com/) benchmark, publicly available set of nine evaluation tasks for the Polish language understanding.
It had the best average performance and obtained the best results for three of them.

| Model | Average | NKJP-NER | CDSC-E | CDSC-R | CBD | PolEmo2.0-IN\t|PolEmo2.0-OUT | DYK | PSC | AR\t|
| :------ | ------: | ------: | ------: | ------: | ------: | ------: | ------: |  ------: | ------: | ------: |
| herbert-klej-cased-v1 | **80.5** | 92.7 | 92.5 | 91.9 | **50.3** | **89.2** |**76.3** |52.1 |95.3 | 84.5 |

Full leaderboard is available [online](https://klejbenchmark.com/leaderboard). 


## HerBERT usage
Model training and experiments were conducted with [transformers](https://github.com/huggingface/transformers) in version 2.0.

Example code:
```python
from transformers import XLMTokenizer, RobertaModel

tokenizer = XLMTokenizer.from_pretrained("allegro/herbert-klej-cased-tokenizer-v1")
model = RobertaModel.from_pretrained("allegro/herbert-klej-cased-v1")

encoded_input = tokenizer.encode("Kto ma lepszą sztukę, ma lepszy rząd – to jasne.", return_tensors='pt')
outputs = model(encoded_input)
```

HerBERT can also be loaded using `AutoTokenizer` and `AutoModel`:

```python
tokenizer = AutoTokenizer.from_pretrained("allegro/herbert-klej-cased-tokenizer-v1")
model = AutoModel.from_pretrained("allegro/herbert-klej-cased-v1")
```

## License
CC BY-SA 4.0

## Citation
If you use this model, please cite the following paper:
```
@inproceedings{rybak-etal-2020-klej,
    title = "{KLEJ}: Comprehensive Benchmark for {P}olish Language Understanding",
    author = "Rybak, Piotr  and
      Mroczkowski, Robert  and
      Tracz, Janusz  and
      Gawlik, Ireneusz",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.111",
    doi = "10.18653/v1/2020.acl-main.111",
    pages = "1191--1201",
}
```

## Authors
The model was trained by **Allegro Machine Learning Research** team.

You can contact us at: <a href="mailto:klejbenchmark@allegro.pl">klejbenchmark@allegro.pl</a>