Edit model card

KB-BART

A BART model trained on a Swedish corpus consisting of 15 billion tokens (about 80GB of text). The model was trained with Fairseq, and converted to be compatible with Huggingface.

Training code can be found here.

Usage

from transformers import BartForConditionalGeneration, PreTrainedTokenizerFast, AutoTokenizer

model = BartForConditionalGeneration.from_pretrained("KBLab/bart-base-swedish-cased")
tok = AutoTokenizer.from_pretrained("KBLab/bart-base-swedish-cased")

model.eval()

input_ids = tok.encode(
    "Jag har ätit en utsökt <mask> på restaurang vid <mask> .", return_tensors="pt"
)

# Simple greedy search
output_ids = model.generate(
    input_ids,
    min_length=15,
    max_length=25,
    num_beams=1,
    do_sample=False,
)
tok.decode(output_ids[0])
# '</s><s> Jag har ätit en utsökt middag på restaurang vid havet på restaurang vid havet på restaurang vid havet.</s>'


# Sampling
output_ids = model.generate(
    input_ids,
    min_length=15,
    max_length=20,
    num_beams=1,
    do_sample=True,
)
tok.decode(output_ids[0])
#'</s><s> Jag har ätit en utsökt god mat som de tagit in på restaurang vid avröjda</s>'


# Beam search
output_ids = model.generate(
    input_ids,
    min_length=15,
    max_length=25,
    no_repeat_ngram_size=3,
    num_beams=8,
    early_stopping=True,
    do_sample=True,
    num_return_sequences=6
)
tok.decode(output_ids[0])
# '</s><s> Jag har ätit en utsökt middag på restaurang vid havet. Jag har varit ute och gått en sväng.</s><pad><pad>'


# Diverse beam generation
output_ids = model.generate(
    input_ids,
    min_length=50,
    max_length=100,
    no_repeat_ngram_size=3,
    num_beams=8,
    early_stopping=True,
    do_sample=False,
    num_return_sequences=6,
    num_beam_groups=8,
    diversity_penalty=2.0,
)
tok.decode(output_ids[0])
# '</s><s> Jag har ätit en utsökt middag på restaurang vid havet på restaurang. Jag har varit på restaurang i två dagar... Jag..,..!!!.. Så.. Nu.. Hej.. Vi.. Här.</s>'

Acknowledgements

We gratefully acknowledge the HPC RIVR consortium (www.hpc-rivr.si) and EuroHPC JU (eurohpc-ju.europa.eu/) for funding this research by providing computing resources of the HPC system Vega at the Institute of Information Science (www.izum.si).

Downloads last month
274
Safetensors
Model size
139M params
Tensor type
F32
·
Inference Examples
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.

Space using KBLab/bart-base-swedish-cased 1

Collection including KBLab/bart-base-swedish-cased