legal_t5_small_trans_cs_sv_small_finetuned model
Model on translating legal text from Cszech to Swedish. It was first released in this repository. This model is first pretrained all the translation data over some unsupervised task. Then the model is trained on three parallel corpus from jrc-acquis, europarl and dcep.
Model description
legal_t5_small_trans_cs_sv_small_finetuned is initially pretrained on unsupervised task with the all of the data of the training set. The unsupervised task was "masked language modelling". legal_t5_small_trans_cs_sv_small_finetuned is based on the t5-small
model and was trained on a large corpus of parallel text. This is a smaller model, which scales the baseline model of t5 down by using dmodel = 512
, dff = 2,048
, 8-headed attention, and only 6 layers each in the encoder and decoder. This variant has about 60 million parameters.
Intended uses & limitations
The model could be used for translation of legal texts from Cszech to Swedish.
How to use
Here is how to use this model to translate legal text from Cszech to Swedish in PyTorch:
from transformers import AutoTokenizer, AutoModelWithLMHead, TranslationPipeline
pipeline = TranslationPipeline(
model=AutoModelWithLMHead.from_pretrained("SEBIS/legal_t5_small_trans_cs_sv_small_finetuned"),
tokenizer=AutoTokenizer.from_pretrained(pretrained_model_name_or_path = "SEBIS/legal_t5_small_trans_cs_sv", do_lower_case=False,
skip_special_tokens=True),
device=0
)
cs_text = "10 Ukončení denního zasedání"
pipeline([cs_text], max_length=512)
Training data
The legal_t5_small_trans_cs_sv_small_finetuned (the supervised task which involved only the corresponding langauge pair and as well as unsupervised task where all of the data of all language pairs were available) model was trained on JRC-ACQUIS, EUROPARL, and DCEP dataset consisting of 5 Million parallel texts.
Training procedure
The model was trained on a single TPU Pod V3-8 for 250K steps in total, using sequence length 512 (batch size 4096). It has a total of approximately 220M parameters and was trained using the encoder-decoder architecture. The optimizer used is AdaFactor with inverse square root learning rate schedule for pre-training.
Preprocessing
An unigram model trained with 88M lines of text from the parallel corpus (of all possible language pairs) to get the vocabulary (with byte pair encoding), which is used with this model.
Pretraining
The pre-training data was the combined data from all the 42 language pairs. The task for the model was to predict the portions of a sentence which were masked randomly.
Evaluation results
When the model is used for translation test dataset, achieves the following results:
Test results :
Model | BLEU score |
---|---|
legal_t5_small_trans_cs_sv_small_finetuned | 48.159 |
BibTeX entry and citation info
Created by Ahmed Elnaggar/@Elnaggar_AI | LinkedIn
- Downloads last month
- 17