By the Hellenic Army Academy (SSE) and the Technical University of Crete (TUC)
language: el
licence: apache-2.0
dataset: ~23.4 GB of Greek corpora
model: GPT2 (12-layer, 768-hidden, 12-heads, 117M parameters. OpenAI GPT-2 English model, finetuned for the Greek language)
pre-processing: tokenization + BPE segmentation
metrics: perplexity
Model description
A text generation (autoregressive) model, using Huggingface transformers and fastai based on the English GPT-2.
Finetuned with gradual layer unfreezing. This is a more efficient and sustainable alternative compared to training from scratch, especially for low-resource languages.
Based on the work of Thomas Dehaene (ML6) for the creation of a Dutch GPT2: https://colab.research.google.com/drive/1Y31tjMkB8TqKKFlZ5OJ9fcMp3p8suvs4?usp=sharing
How to use
from transformers import pipeline
model = "lighteternal/gpt2-finetuned-greek"
generator = pipeline(
'text-generation',
device=0,
model=f'{model}',
tokenizer=f'{model}')
text = "Μια φορά κι έναν καιρό"
print("\
".join([x.get("generated_text") for x in generator(
text,
max_length=len(text.split(" "))+15,
do_sample=True,
top_k=50,
repetition_penalty = 1.2,
add_special_tokens=False,
num_return_sequences=5,
temperature=0.95,
top_p=0.95)]))
Training data
We used a 23.4GB sample from a consolidated Greek corpus from CC100, Wikimatrix, Tatoeba, Books, SETIMES and GlobalVoices containing long senquences.
This is a better version of our GPT-2 small model (https://huggingface.co/lighteternal/gpt2-finetuned-greek-small)
Metrics
Metric
Value
Train Loss
3.67
Validation Loss
3.83
Perplexity
39.12
Acknowledgement
The research work was supported by the Hellenic Foundation for Research and Innovation (HFRI) under the HFRI PhD Fellowship grant (Fellowship Number:50, 2nd call)
