Команда:

MERA

Название ML-модели:

ruGPT-3-large

Ссылка на ML-модель:

Дополнительные ссылки:

Описание архитектуры:

ruGPT-3 is a Russian counterpart of GPT-3 (Brown et al., 2020). We use the model architecture description by Brown et al. and the GPT-2 code base (Radford et al., 2019) from the Transformers library. ruGPT-3 is pretrained on the language modeling objective. The BBPE tokenizer with the vocabulary size of 5 · 104 tokens was used.

Описание обучения:

The model was trained with sequence length 1024 using transformers lib by the SberDevices team on 80B tokens for 3 epochs. After that, the model was finetuned 1 epoch with sequence length 2048. Total training time was around 14 days on 128 GPUs for 1024 context and a few days on 16 GPUs for 2048 context. The final perplexity on the test set is 13.6.

Данные претрейна:

450GB of texts. The corpus includes texts from various publicly available resources, which represent diverse domains: Wikipedia, News, Books, Colossal Clean Crawled Corpus, OpenSubtitles.

Детали обучения:

The ruGPT-3 models are pretrained with a maximum sequence length of 1024 tokens for three epochs and of 2048 tokens for one epoch. We use the initial learning rate of 1e−4 and the Adam optimizer with β1 = 0.9, β2 = 0.99, and ϵ = 1e−8. The total number of tokens seen during pretraining is 80B. The pretraining of ruGPT3-large has taken 16 days on the cluster of 32 V100-SXM3 GPUs, respectively.

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 2048