README.md

---
language:
- ru
license: apache-2.0
---

# FRED-T5 1.7B (Full-scale Russian Enhanced Denoisers T5) 

Architecture based on T5. 

It has 24 layers and 1536 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/1936

Total training time was around 45 days on 112 A100 GPUs.

Training loss
![Screenshot 2023-01-21 at 11.36.52.png](https://s3.amazonaws.com/moonup/production/uploads/1674290304538-5f91b1208a61a359f44e1851.png)

We continue to experiment... 

We'll  release  checkpoint  to the public soon.