|
--- |
|
language: |
|
- sl |
|
|
|
license: cc-by-sa-4.0 |
|
--- |
|
# Usage |
|
Load in transformers library with: |
|
``` |
|
from transformers import AutoTokenizer, AutoModelForMaskedLM |
|
|
|
tokenizer = AutoTokenizer.from_pretrained("EMBEDDIA/sloberta", use_fast=False) |
|
model = AutoModelForMaskedLM.from_pretrained("EMBEDDIA/sloberta") |
|
``` |
|
**NOTE**: it is currently *critically important* to add `use_fast=False` parameter to tokenizer if using transformers version 4+ (prior versions have `use_fast=False` as default) By default it attempts to load a fast tokenizer, which will work (ie. not result in an error), but it will not correctly map tokens to its IDs and the performance on any task will be extremely bad. |
|
|
|
# SloBERTa |
|
SloBERTa model is a monolingual Slovene BERT-like model. It is closely related to French Camembert model https://camembert-model.fr/. The corpora used for training the model have 3.47 billion tokens in total. The subword vocabulary contains 32,000 tokens. The scripts and programs used for data preparation and training the model are available on https://github.com/clarinsi/Slovene-BERT-Tool |
|
|
|
SloBERTa was trained for 200,000 iterations or about 98 epochs. |
|
|
|
## Corpora |
|
The following corpora were used for training the model: |
|
* Gigafida 2.0 |
|
* Kas 1.0 |
|
* Janes 1.0 (only Janes-news, Janes-forum, Janes-blog, Janes-wiki subcorpora) |
|
* Slovenian parliamentary corpus siParl 2.0 |
|
* slWaC |
|
|
|
|