Команда:

MERA

Название ML-модели:

FRED-T5 large 820M

Ссылка на ML-модель:

Дополнительные ссылки:

Описание архитектуры:

FRED-T5 (Full-scale Russian Enhanced Denoisers) is an encoder-decoder model based on T5 and UL2. Number of attantion heads 16. The dimensions of the hidden layers 1024 and the fully connected layers 2816. GELU activation function.

Описание обучения:

Bbpe tokenizer. 50257 + special tokens 107. Prefix tokens: '<LM>', '<SC1>',.. '<SC6>'. Drawing inspiration from Tay et al. (2022), the FRED-T5 1.7.B (or XL) model was pretrained on a mixture of denoisers (MoD), a pretraining objective that represents a set of diverse pretraining objectives. The R-Denoiser is a masked language modeling span corruption objective used in T5. The S-Denoiser follows the language modeling objective, where the input sequence is split into the context and target tokens so that the targets do not rely on future information. The X-Denoiser aims to recover a large part of the input based on the span corruption and language modeling objectives.

Данные претрейна:

It was trained on Russian language corpus (300GB).

Детали обучения:

FRED-T5 is pretrained using a linear scheduler with the initial learning rate of 1e−4 and the Adafactor optimizer (Shazeer and Stern, 2018) with β1 = 0.9, β2 = 0.99, and ϵ = 1e−8. The sequence length is set to 512/512 for inputs and targets. The FRED-T5-XL models is pretrained pretrained with a total batch size of 2048 for 35 days on 160 V100 GPUs, followed by 5 days on 80 A100 GPUs.

Лицензия:

MIT

Стратегия, генерация и параметры:

Code version v.1.1.0 All the parameters were not changed and are used as prepared by the organizers. Details: - 1 x NVIDIA A100 - dtype auto - Pytorch 2.1.2 + CUDA 12.1 - Transformers 4.36.2 - Context length 512