cointegrated/rut5-small

🤗 Hugging Face sourcemit65M params259 MBsafetensors✓ 4 checksumsupdated today
Submit in one command

Run it next to your model folder. It makes the torrent, checks your files against Hugging Face, and submits it. You just start seeding and paste your key from your account. It only reads your files and never changes them. Read the script first if you like.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo cointegrated/rut5-small ./model-folder
Needs a seeder →

This is a small Russian paraphraser based on the google/mt5-small model. It has rather poor paraphrasing performance, but can be fine tuned for this or other tasks.

This model was created by taking the alenusch/mt5small-ruparaphraser model and stripping 96% of its vocabulary which is unrelated to the Russian language or infrequent.

  • The original model has 300M parameters, with 256M of them being input and output embeddings.
  • After shrinking the sentencepiece vocabulary from 250K to 20K the number of model parameters reduced to 65M parameters, and model size reduced from 1.1GB to 246MB.
    • The first 5K tokens in the new vocabulary are taken from the original mt5-small.
    • The next 15K tokens are the most frequent tokens obtained by tokenizing a Russian web corpus from the Leipzig corpora collection.

The model can be used as follows:

# !pip install transformers sentencepiece
import torch
from transformers import T5ForConditionalGeneration, T5Tokenizer

tokenizer = T5Tokenizer.from_pretrained("cointegrated/rut5-small")
model = T5ForConditionalGeneration.from_pretrained("cointegrated/rut5-small")

text = 'Ехал Грека через реку, видит Грека в реке рак. '
inputs = tokenizer(text, return_tensors='pt')
with torch.no_grad():
    hypotheses = model.generate(
        **inputs, 
        do_sample=True, top_p=0.95, num_return_sequences=10, 
        repetition_penalty=2.5,
        max_length=32,
    )
for h in hypotheses:
    print(tokenizer.decode(h, skip_special_tokens=True))