File size: 2,658 Bytes
80817eb
eeccd0e
80817eb
 
 
 
eeccd0e
80817eb
 
 
eeccd0e
80817eb
 
 
 
 
d15358c
 
 
 
 
 
 
 
 
 
 
 
 
80817eb
 
 
 
 
d15358c
 
 
 
80817eb
 
d15358c
 
 
80817eb
 
 
d15358c
80817eb
23ec421
 
 
 
f428795
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
---
license: mit
language:
- fr
library_name: transformers
inference: false
pipeline_tag: feature-extraction
---
# CamemBERT-L4

This model is a pruned version of the pre-trained [CamemBERT](https://huggingface.co/camembert-base) checkpoint, obtained by [dropping the top-layers](https://doi.org/10.48550/arXiv.2004.03844) from the original model.

![](illustration.jpeg)

## Usage

You can use the raw model for masked language modeling (MLM), but it's mostly intended to be fine-tuned on a downstream task, especially one that uses the whole sentence to make decisions such as text classification, extractive question answering, or semantic search. For tasks such as text generation, you should look at autoregressive models like [BelGPT-2](https://huggingface.co/antoinelouis/belgpt2).

You can use this model directly with a pipeline for [masked language modeling](https://huggingface.co/tasks/fill-mask):

```python
from transformers import pipeline

unmasker = pipeline('fill-mask', model='antoinelouis/camembert-L4')
unmasker("Bonjour, je suis un [MASK] modèle.")
```

You can also use this model to [extract the features](https://huggingface.co/tasks/feature-extraction) of a given text:

```python
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained('antoinelouis/camembert-L4')
model = AutoModel.from_pretrained('antoinelouis/camembert-L4')

text = "Remplacez-moi par le texte de votre choix."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```

## Variations

CamemBERT has originally been released in base (110M) and large (335M) variations. The following checkpoints prune the base variation by dropping the top 2, 4, 6, 8, and 10 pretrained encoding layers, respectively.

| Model                                                              | #Params |  Size | Pruning |
|--------------------------------------------------------------------|:-------:|:-----:|:-------:|
| [CamemBERT-base](https://huggingface.co/camembert-base)            |  110.6M | 445MB |    -    |
|                                                                    |         |       |         |
| [CamemBERT-L10](https://huggingface.co/antoinelouis/camembert-L10) |  96.4M  | 386MB |   -13%  |
| [CamemBERT-L8](https://huggingface.co/antoinelouis/camembert-L8)   |  82.3M  | 329MB |   -26%  |
| [CamemBERT-L6](https://huggingface.co/antoinelouis/camembert-L6)   |  68.1M  | 272MB |   -38%  |
| **CamemBERT-L4**                                                   |  53.9M  | 216MB |   -51%  |
| [CamemBERT-L2](https://huggingface.co/antoinelouis/camembert-L2)   |  39.7M  | 159MB |   -64%  |