metadata

library_name: transformers
language:
  - ru
metrics:
  - rouge
base_model:
  - ai-forever/ruT5-base
pipeline_tag: summarization

ruT5-base_headline_generation

Model Details

T5 Base for news headline generation (Russian). The model is finetuned for best performance on short news texts (128 words or less), but it has decent metrics on longer articles as well. The model generates abstractive headlines that on average include 6-11 words.

Base Model: ai-forever/ruT5-base

Training Details

Training Data: 247 000 news articles in Russian

Training Procedure: 6 epochs, all details and hyperparameters in this Google Colab notebook

Testing Metrics

Rouge1: 40.24
Rouge2: 23.05
RougeL: 37.57

How to Use

from transformers import AutoTokenizer, T5ForConditionalGeneration

model_name = "wanderer-msk/ruT5-base_headline_generation"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = T5ForConditionalGeneration.from_pretrained(model_name)

news_text = """Земляне продолжают осваивать Марс.
Колонисты уже посадили на красной планете 42 яблони."""

model_input = tokenizer(
    news_text,
    truncation=True,
    max_length=1024,
    return_tensors="pt"
)
model_output = model.generate(model_input["input_ids"])
news_headline = tokenizer.decode(
    model_output.squeeze(),
    skip_special_tokens=True
)

print(news_headline)