Aeoinum v1 BaseWeb 1B

A state-of-the-art language model for Russian language processing. This checkpoint contains a preliminary version of the model with 1.6 billion parameters. Trained only on web pages.

Models

Name	N of parameters	N of dataset tokens	Context window
Aeonium-v1-BaseWeb-1B	1.6B	32B	4K
Aeonium-v1-Base-1B	1.6B	In training	4K
Aeonium-v1-Chat-1B	1.6B	In training	4K
Aeonium-v1-Base-4B	1.6B	In training	4K
Aeonium-v1-Chat-4B	1.6B	In training	4K

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("aeonium/Aeonium-v1-BaseWeb-1B")
model = AutoModelForCausalLM.from_pretrained("aeonium/Aeonium-v1-BaseWeb-1B").cuda()

input_ids = tokenizer("Искусственный интеллект - это", return_tensors='pt').to(model.device)["input_ids"]
output = model.generate(input_ids, max_new_tokens=48, do_sample=True, temperature=0.7)
print(tokenizer.decode(output[0]))

Output:

Искусственный интеллект - это не только про компьютеры и смартфоны. Его возможности безграничны, а с развитием интернета и интернета вещей он становится еще и самым настоящим оружием в борьбе с преступностью.
Мы поговорили с юристом о самых интересных и опасных способах использования ИИ.

Dataset Detail

The dataset for pre-training is collected from public data, most of which are web pages in Russian. The total size of the data is 32B tokens.

Training Detail

The training is performed thanks to a grant from TPU Research Cloud on a TPU v4-128 node.

Loss: 2.68; Accuracy: 0.48; Batch Size: 1024

Content Warning

Aeonium v1 is a large language model trained on a broad dataset from the internet. As such, it may generate text that contains biases, offensive language, or other disapproving content. The model outputs should not be considered factual or representative of any individual's beliefs or identity. Users should exercise caution and apply careful filtering when using Aeonium's generated text, especially for sensitive or high-stakes applications. The developers do not condone generating harmful, biased, or unethical content.

Copyright

The model is released under the Apache 2.0 license.