ModernBERT-JP-0.5B-PT-stage2 builds on iamtatsuki05/ModernBERT-JP-0.5B-PT-stage1 with additional pre-training on fujiki/wiki40b_ja. The model observes roughly 1B tokens under 8,192-token contexts to improve encyclopedic coverage and long-context understanding in Japanese.

Hugging Face Collection
GitHub

Usage

Requirements

transformers>=4.51.0
accelerate>=1.6.0
sentencepiece>=0.2.0
flash-attn>=2.7.3

Sample Code

python

import torch
from transformers import AutoModelForMaskedLM, AutoTokenizer

model_name = "iamtatsuki05/ModernBERT-JP-0.5B-PT-stage2"
model_kwargs = {
  "torch_dtype": torch.bfloat16,
  "attn_implementation": "flash_attention_2",
  "device_map": "auto",
}
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, **model_kwargs)

text = f"ハチワレは{tokenizer.mask_token}のキャラクターです。"
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model(**inputs)
masked_index = inputs["input_ids"][0].tolist().index(tokenizer.mask_token_id)
print(tokenizer.decode(outputs.logits[0, masked_index].argmax(axis=-1)))

Model Details

Base model: iamtatsuki05/ModernBERT-JP-0.5B-PT-stage1
Architecture: ModernBERT
Maximum sequence length: 8,192 tokens
Embedding dimension: 1280
Tokenizer: SentencePiece / vocabulary size 102,400
Positional encoding: RoPE
Supported languages: Japanese

Model Series

These models are further pre-trained on fujiki/wiki40b_ja for roughly 1B tokens with 8,192-token sequence lengths to enrich encyclopedic knowledge.

ID	Architecture	#Param.	#Param. w/o Emb.	JGLUE-Avg	JGLUE-JSTS	JGLUE-JNLI	JGLUE-JCoLA
iamtatsuki05/ModernBERT-JP-0.5B-PT-stage2 (this model)	ModernBERT	679M	548M	86.32	89.29	86.68	83.00
iamtatsuki05/Llama-JP-0.5B-PT-stage2	Llama	661M	530M	81.88	82.91	78.67	84.06

Licence

This model is distributed under the MIT License.

How to Cite

tex

@article{MIREI
  title={同一条件下における Encoder/Decoder アーキテクチャによる文埋め込みの性能分析},
  author={岡田 龍樹 and 杉本 徹},
  journal={言語処理学会第 32 回年次大会 (NLP2026)},
  year={2026}
}