license: mit
language:
- fr
library_name: transformers
inference: false
pipeline_tag: feature-extraction
CamemBERT-L2
This model is a pruned version of the pre-trained CamemBERT checkpoint, obtained by dropping the top-layers from the original model.
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.
You can use this model directly with a pipeline for masked language modeling:
from transformers import pipeline
unmasker = pipeline('fill-mask', model='antoinelouis/camembert-L2')
unmasker("Bonjour, je suis un [MASK] modèle.")
You can also use this model to extract the features of a given text:
from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained('antoinelouis/camembert-L2')
model = AutoModel.from_pretrained('antoinelouis/camembert-L2')
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 | 110.6M | 445MB | - |
| CamemBERT-L10 | 96.4M | 386MB | -13% |
| CamemBERT-L8 | 82.3M | 329MB | -26% |
| CamemBERT-L6 | 68.1M | 272MB | -38% |
| CamemBERT-L4 | 53.9M | 216MB | -51% |
| CamemBERT-L2 | 39.7M | 159MB | -64% |