camembert-base / README.md
1 ---
2 language: fr
3 license: mit
4 datasets:
5 - oscar
6 ---
7
8 # CamemBERT: a Tasty French Language Model
9
10 ## Introduction
11
12 [CamemBERT](https://arxiv.org/abs/1911.03894) is a state-of-the-art language model for French based on the RoBERTa model.
13
14 It is now available on Hugging Face in 6 different versions with varying number of parameters, amount of pretraining data and pretraining data source domains.
15
16 For further information or requests, please go to [Camembert Website](https://camembert-model.fr/)
17
18 ## Pre-trained models
19
20 | Model | #params | Arch. | Training data |
21 |--------------------------------|--------------------------------|-------|-----------------------------------|
22 | `camembert-base` | 110M | Base | OSCAR (138 GB of text) |
23 | `camembert/camembert-large` | 335M | Large | CCNet (135 GB of text) |
24 | `camembert/camembert-base-ccnet` | 110M | Base | CCNet (135 GB of text) |
25 | `camembert/camembert-base-wikipedia-4gb` | 110M | Base | Wikipedia (4 GB of text) |
26 | `camembert/camembert-base-oscar-4gb` | 110M | Base | Subsample of OSCAR (4 GB of text) |
27 | `camembert/camembert-base-ccnet-4gb` | 110M | Base | Subsample of CCNet (4 GB of text) |
28
29 ## How to use CamemBERT with HuggingFace
30
31 ##### Load CamemBERT and its sub-word tokenizer :
32 ```python
33 from transformers import CamembertModel, CamembertTokenizer
34
35 # You can replace "camembert-base" with any other model from the table, e.g. "camembert/camembert-large".
36 tokenizer = CamembertTokenizer.from_pretrained("camembert-base")
37 camembert = CamembertModel.from_pretrained("camembert-base")
38
39 camembert.eval() # disable dropout (or leave in train mode to finetune)
40
41 ```
42
43 ##### Filling masks using pipeline
44 ```python
45 from transformers import pipeline
46
47 camembert_fill_mask = pipeline("fill-mask", model="camembert-base", tokenizer="camembert-base")
48 results = camembert_fill_mask("Le camembert est <mask> :)")
49 # results
50 #[{'sequence': '<s> Le camembert est délicieux :)</s>', 'score': 0.4909103214740753, 'token': 7200},
51 # {'sequence': '<s> Le camembert est excellent :)</s>', 'score': 0.10556930303573608, 'token': 2183},
52 # {'sequence': '<s> Le camembert est succulent :)</s>', 'score': 0.03453315049409866, 'token': 26202},
53 # {'sequence': '<s> Le camembert est meilleur :)</s>', 'score': 0.03303130343556404, 'token': 528},
54 # {'sequence': '<s> Le camembert est parfait :)</s>', 'score': 0.030076518654823303, 'token': 1654}]
55
56 ```
57
58 ##### Extract contextual embedding features from Camembert output
59 ```python
60 import torch
61 # Tokenize in sub-words with SentencePiece
62 tokenized_sentence = tokenizer.tokenize("J'aime le camembert !")
63 # ['▁J', "'", 'aime', '▁le', '▁ca', 'member', 't', '▁!']
64
65 # 1-hot encode and add special starting and end tokens
66 encoded_sentence = tokenizer.encode(tokenized_sentence)
67 # [5, 121, 11, 660, 16, 730, 25543, 110, 83, 6]
68 # NB: Can be done in one step : tokenize.encode("J'aime le camembert !")
69
70 # Feed tokens to Camembert as a torch tensor (batch dim 1)
71 encoded_sentence = torch.tensor(encoded_sentence).unsqueeze(0)
72 embeddings, _ = camembert(encoded_sentence)
73 # embeddings.detach()
74 # embeddings.size torch.Size([1, 10, 768])
75 # tensor([[[-0.0254, 0.0235, 0.1027, ..., -0.1459, -0.0205, -0.0116],
76 # [ 0.0606, -0.1811, -0.0418, ..., -0.1815, 0.0880, -0.0766],
77 # [-0.1561, -0.1127, 0.2687, ..., -0.0648, 0.0249, 0.0446],
78 # ...,
79 ```
80
81 ##### Extract contextual embedding features from all Camembert layers
82 ```python
83 from transformers import CamembertConfig
84 # (Need to reload the model with new config)
85 config = CamembertConfig.from_pretrained("camembert-base", output_hidden_states=True)
86 camembert = CamembertModel.from_pretrained("camembert-base", config=config)
87
88 embeddings, _, all_layer_embeddings = camembert(encoded_sentence)
89 # all_layer_embeddings list of len(all_layer_embeddings) == 13 (input embedding layer + 12 self attention layers)
90 all_layer_embeddings[5]
91 # layer 5 contextual embedding : size torch.Size([1, 10, 768])
92 #tensor([[[-0.0032, 0.0075, 0.0040, ..., -0.0025, -0.0178, -0.0210],
93 # [-0.0996, -0.1474, 0.1057, ..., -0.0278, 0.1690, -0.2982],
94 # [ 0.0557, -0.0588, 0.0547, ..., -0.0726, -0.0867, 0.0699],
95 # ...,
96 ```
97
98
99 ## Authors
100
101 CamemBERT was trained and evaluated by Louis Martin\*, Benjamin Muller\*, Pedro Javier Ortiz Suárez\*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
102
103
104 ## Citation
105 If you use our work, please cite:
106
107 ```bibtex
108 @inproceedings{martin2020camembert,
109 title={CamemBERT: a Tasty French Language Model},
110 author={Martin, Louis and Muller, Benjamin and Su{\'a}rez, Pedro Javier Ortiz and Dupont, Yoann and Romary, Laurent and de la Clergerie, {\'E}ric Villemonte and Seddah, Djam{\'e} and Sagot, Beno{\^\i}t},
111 booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
112 year={2020}
113 }
114 ```
115
116