racai/distilbert-base-romanian-uncased

Romanian DistilBERT

This repository contains the uncased Romanian DistilBERT (named Distil-RoBERT-base in the paper). The teacher model used for distillation is: readerbench/RoBERT-base.

The model was introduced in this paper. The adjacent code can be found here.

Usage

from transformers import AutoTokenizer, AutoModel

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained("racai/distilbert-base-romanian-uncased")
model = AutoModel.from_pretrained("racai/distilbert-base-romanian-uncased")

# tokenize a test sentence
input_ids = tokenizer.encode("aceasta este o propoziție de test.", add_special_tokens=True, return_tensors="pt")

# run the tokens trough the model
outputs = model(input_ids)

print(outputs)

Model Size

It is 35% smaller than its teacher RoBERT-base.

Model	Size (MB)	Params (Millions)
RoBERT-base	441	114
distilbert-base-romanian-cased	282	72

Evaluation

We evaluated the model in comparison with the RoBERT-base on 5 Romanian tasks:

• UPOS: Universal Part of Speech (F1-macro)
• XPOS: Extended Part of Speech (F1-macro)
• NER: Named Entity Recognition (F1-macro)
• SAPN: Sentiment Anlaysis - Positive vs Negative (Accuracy)
• SAR: Sentiment Analysis - Rating (F1-macro)
• DI: Dialect identification (F1-macro)
• STS: Semantic Textual Similarity (Pearson)

Model	UPOS	XPOS	NER	SAPN	SAR	DI	STS
RoBERT-base	98.02	97.15	85.14	98.30	79.40	96.07	81.18
distilbert-base-romanian-uncased	97.12	95.79	83.11	98.01	79.58	96.11	79.80

BibTeX entry and citation info

@article{avram2021distilling,
  title={Distilling the Knowledge of Romanian BERTs Using Multiple Teachers},
  author={Andrei-Marius Avram and Darius Catrina and Dumitru-Clementin Cercel and Mihai Dascălu and Traian Rebedea and Vasile Păiş and Dan Tufiş},
  journal={ArXiv},
  year={2021},
  volume={abs/2112.12650}
}