import torch
from transformers import AutoTokenizer, AutoModelForMaskedLM

# Import model
tokenizer = AutoTokenizer.from_pretrained(
    "HPLT/hplt_gpt_bert_base_3_0_rus_Cyrl",
)
model = AutoModelForMaskedLM.from_pretrained(
    "HPLT/hplt_gpt_bert_base_3_0_rus_Cyrl",
    trust_remote_code=True,
    use_safetensors=False,
)
model = model.eval()
input_text = f"Я шёл по тропинке с двумя друзьями — {tokenizer.mask_token} садилось"
print(input_text)
# Tokenize text (with a mask token inside)
input_text = tokenizer(
    input_text,
    return_tensors="pt",
)
# Inference
with torch.no_grad():
    output_p = model(**input_text)

# Unmask the text
output_text = torch.where(
    input_text.input_ids == tokenizer.mask_token_id,
    output_p.logits.argmax(-1),
    input_text.input_ids
)

# Decoding; should output: '[CLS]Я шёл по тропинке с двумя друзьями — солнце садилось[SEP]'
print(tokenizer.decode(output_text[0].tolist()))

Example usage (text generation)

GPT-BERT also supports unidirectional text decoding, it can generate text like any other GPT model:

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained(
    "HPLT/hplt_gpt_bert_base_3_0_rus_Cyrl",
)
model = AutoModelForCausalLM.from_pretrained(
    "HPLT/hplt_gpt_bert_base_3_0_rus_Cyrl",
    trust_remote_code=True, 
    use_safetensors=False,
)
text = f"Я шёл по тропинке с двумя друзьями"
print(text, flush=True)
# Define tokens that should end the generation
eos_token_ids = [
    token_id
    for token_id in range(tokenizer.vocab_size)
    if '.' in tokenizer.decode([token_id])
]

# Generation function
@torch.no_grad()
def generate(text):
    input_ids = tokenizer(text, return_tensors='pt').input_ids
    prediction = model.generate(
        input_ids,
        max_new_tokens=63,
        do_sample=False,
        eos_token_id=eos_token_ids,
    )
    return tokenizer.decode(prediction[0]).strip()

# Example usage, should output '[CLS]Я шёл по тропинке с двумя друзьями[SEP] и увидел на горизонте большую гору с двумя высокими холмами.'
print(generate(text), flush=True)

The following classes are currently implemented: AutoModel, AutoModelForMaskedLM, AutoModelForCausalLM, AutoModelForSequenceClassification, AutoModelForTokenClassification, AutoModelForQuestionAnswering and AutoModeltForMultipleChoice.

Intermediate checkpoints

We are releasing 10 intermediate checkpoints for each model at intervals of every 3125 training steps in separate branches. The naming convention is stepXXX: for example, step18750.

You can load a specific model revision with transformers using the argument revision:

model = AutoModelForSeq2SeqLM.from_pretrained("HPLT/hplt_gpt_bert_base_3_0_rus_Cyrl", revision="step21875", trust_remote_code=True)

You can access all the revisions for the models with the following code:

python

from huggingface_hub import list_repo_refs
out = list_repo_refs("HPLT/hplt_gpt_bert_base_3_0_rus_Cyrl")
print([b.name for b in out.branches])

Cite us

bibtex

@inproceedings{charpentier-samuel-2024-bert,
    title = "{GPT} or {BERT}: why not both?",
    author = "Charpentier, Lucas Georges Gabriel  and
      Samuel, David",
    booktitle = "The 2nd BabyLM Challenge at the 28th Conference on Computational Natural Language Learning",
    month = nov,
    year = "2024",
    address = "Miami, FL, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.conll-babylm.24/",
    pages = "262--283"
}

bibtex

@misc{oepen2025hplt30largescalemultilingual,
      title={{HPLT 3.0}: {V}ery Large-Scale Multilingual Resources for {LLM} and {MT}. Mono- and Bi-lingual Data, Multilingual Evaluation, and Pre-Trained Models}, 
      author={Stephan Oepen and Nikolay Arefev and Mikko Aulamo and Marta Bañón and Maja Buljan and Laurie Burchell and Lucas Charpentier and Pinzhen Chen and Mariia Fedorova and Ona de Gibert and Barry Haddow and Jan Hajič and Jindřich Helcl and Andrey Kutuzov and Veronika Laippala and Zihao Li and Risto Luukkonen and Bhavitvya Malik and Vladislav Mikhailov and Amanda Myntti and Dayyán O'Brien and Lucie Poláková and Sampo Pyysalo and Gema Ramírez Sánchez and Janine Siewert and Pavel Stepachev and Jörg Tiedemann and Teemu Vahtola and Dušan Variš and Fedor Vitiugin and Tea Vojtěchová and Jaume Zaragoza},
      year={2025},
      eprint={2511.01066},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.01066}, 
}

arXiv

This project has received funding from the European Union’s Horizon Europe research and innovation programme under grant agreement No 101070350 and from UK Research and Innovation (UKRI) under the UK government’s Horizon Europe funding guarantee [grant number 10052546].