mBART ERRnews
Pretrained mbart-large-cc25 model finetuned on ERRnews Estonian news story dataset.
How to use
Here is how to use this model to get a summary of a given text in PyTorch:
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("TalTechNLP/mBART-ERRnews")
model = AutoModelForSeq2SeqLM.from_pretrained("TalTechNLP/mBART-ERRnews")
text = "Riigikogu rahanduskomisjon võttis esmaspäeval maha riigieelarvesse esitatud investeeringuettepanekutest siseministeeriumi investeeringud koolidele ja lasteaedadele, sest komisjoni hinnangul ei peaks siseministeerium tegelema investeeringutega väljaspoole oma vastutusala. Komisjoni esimees Aivar Kokk ütles, et komisjon lähtus otsuse tegemisel riigikontrolör Janar Holmi soovitusest ja seadustest."
inputs = tokenizer(text, return_tensors='pt', max_length=1024)
summary_ids = model.generate(inputs['input_ids'])
summary = [tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids]
Training data
The mBART model was finetuned on ERRnews, a dataset consisting of 10 420 Estonian news story transcripts and summaries.
Training
The model was trained on 2 cloud GPUs with a batch size of 16 for 16 epochs. The optimizer used is Adam with a learning rate of 5e-05, betas of 0.9 and 0.999.
Evaluation results
This model achieves the following results:
Dataset | ROUGE-1 | ROUGE-2 | ROUGE-L | ROUGE-L-SUM |
---|---|---|---|---|
ERRnews | 19.2 | 6.7 | 16.1 | 17.4 |
BibTeX entry and citation info
article{henryabstractive,
title={Abstractive Summarization of Broadcast News Stories for {Estonian}},
author={Henry, H{\"a}rm and Tanel, Alum{\"a}e},
journal={Baltic J. Modern Computing},
volume={10},
number={3},
pages={511-524},
year={2022}
}
- Downloads last month
- 10
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.