RuGPT3MediumSumGazeta

Model description

This is the model for abstractive summarization for Russian based on rugpt3medium_based_on_gpt2.

Intended uses & limitations

How to use

Colab: link

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "IlyaGusev/rugpt3medium_sum_gazeta"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

article_text = "..."

text_tokens = tokenizer(
    article_text,
    max_length=600,
    add_special_tokens=False, 
    padding=False,
    truncation=True
)["input_ids"]
input_ids = text_tokens + [tokenizer.sep_token_id]
input_ids = torch.LongTensor([input_ids])

output_ids = model.generate(
    input_ids=input_ids,
    no_repeat_ngram_size=4
)

summary = tokenizer.decode(output_ids[0], skip_special_tokens=False)
summary = summary.split(tokenizer.sep_token)[1]
summary = summary.split(tokenizer.eos_token)[0]
print(summary)

Training data

Training procedure

Eval results

  • Train dataset: Gazeta v1 train
  • Test dataset: Gazeta v1 test
  • Source max_length: 600
  • Target max_length: 200
  • no_repeat_ngram_size: 4
  • num_beams: 5
Model R-1-f R-2-f R-L-f chrF METEOR BLEU Avg char length
mbart_ru_sum_gazeta 32.4 14.3 28.0 39.7 26.4 12.1 371
rut5_base_sum_gazeta 32.2 14.4 28.1 39.8 25.7 12.3 330
rugpt3medium_sum_gazeta 26.2 7.7 21.7 33.8 18.2 4.3 244
  • Train dataset: Gazeta v1 train
  • Test dataset: Gazeta v2 test
  • Source max_length: 600
  • Target max_length: 200
  • no_repeat_ngram_size: 4
  • num_beams: 5
Model R-1-f R-2-f R-L-f chrF METEOR BLEU Avg char length
mbart_ru_sum_gazeta 28.7 11.1 24.4 37.3 22.7 9.4 373
rut5_base_sum_gazeta 28.6 11.1 24.5 37.2 22.0 9.4 331
rugpt3medium_sum_gazeta 24.1 6.5 19.8 32.1 16.3 3.6 242

Evaluation script: evaluate.py

Flags: --language ru --tokenize-after --lower

New

Select AutoNLP in the “Train” menu to fine-tune this model automatically.

Downloads last month
79
Hosted inference API
Summarization

Inference API has been turned off for this model.