FRED-T5 large 820M (Full-scale Russian Enhanced Denoisers T5)

Model was trained by SberDevices .

Architecture based on T5.

It has 24 layers and 1024 hidden size. More details in config.json.

The model trained on a mixture of 7 denoisers like UL2 with several differences ( https://arxiv.org/abs/2205.05131 ).

It was trained on Russian language corpus (300GB). The dataset is the same as for ruT5 models.

Bbpe tokenizer. 50257 + special tokens 107. Prefix tokens: '<LM>', '<SC1>',.. '<SC6>'

First half of the time model trained on the small part of all dataset (1%,3GB) and without prefixes in each task.

For RSG, we trained as described in the T5 paper. First, we trained multitask for all tasks. Then we took the best checkpoint for the task and trained it further. RSG submit here https://russiansuperglue.com/login/submit_info/2060

Total training time was around 35 days on 160 V100 GPUs + 5 days on 80 A100.

Usage (HuggingFace Models Repository)

import torch
from transformers import GPT2Tokenizer, T5ForConditionalGeneration 
tokenizer = GPT2Tokenizer.from_pretrained('ai-forever/FRED-T5-1.7B',eos_token='</s>')
model = T5ForConditionalGeneration.from_pretrained('ai-forever/FRED-T5-1.7B')
device='cuda'
model.to(device)

#Prefix <LM>
lm_text='<LM>Принялся Кутузов рассказывать свою историю как он сюда попал. Началось'
input_ids=torch.tensor([tokenizer.encode(lm_text)]).to(device)
outputs=model.generate(input_ids,eos_token_id=tokenizer.eos_token_id,early_stopping=True)
print(tokenizer.decode(outputs[0][1:]))

# print result: , как водится, с того, что он был в плену.</s>

#Prefix <SC1>
lm_text='<SC1>Принялся Кутузов рассказывать свою историю <extra_id_0>. Началось с того, что он был в армии, служил в артиллерии.'
input_ids=torch.tensor([tokenizer.encode(lm_text)]).to(device)
outputs=model.generate(input_ids,eos_token_id=tokenizer.eos_token_id,early_stopping=True)
print(tokenizer.decode(outputs[0][1:]))

#print result: '<extra_id_0>, как он жил</s>'

# Prefix <SC5>
lm_text='<SC5>Принялся Кутузов рассказывать свою историю <extra_id_0>. Началось с того, что он был в армии, служил в артиллерии.'
input_ids=torch.tensor([tokenizer.encode(lm_text)]).to(device)
outputs=model.generate(input_ids,eos_token_id=tokenizer.eos_token_id,early_stopping=True,max_length=100)
print(tokenizer.decode(outputs[0][1:]))

#print result: '<extra_id_0> </s>'

Authors

• NLP core team RnD Telegram channel :
  • Dmitry Zmitrovich
  • Andrei Kalmykov
  • Vitaly Kadulin
  • Mikhail Novikov
  • Alexey Khoroshilov

Salute AI Community .