A Japanese‑language BERT large model (24 layers, 1024 hidden size, 16 heads) pretrained on CC‑100 and Wikipedia using Unidic 2.1.2 word‑level tokenization followed by character splitting, with whole‑word masking for MLM, available via cl-tohoku/bert-japanese.
Description in other languages
Statistics
Statistics as of Nov 16, 2026
Hugging Face downloads (last 30 days)
2,000
Hugging Face likes
2
Listed
May 13, 2024
Popular in the same category
- bert-base-japanese-char-v3tohoku-nlp130kJapanese‑language BERT‑Base (12 layers, 768‑dim, 12 heads) pretrained with Unidic‑based word‑level plus character‑level tokenization and whole‑word masking on CC‑100 and 2023 Wikipedia, producing a 7,027‑token vocabulary.
- bert-large-japanese-v2tohoku-nlp66kJapanese‑BERT‑Large trained on CC‑100 and Wikipedia, using Unidic‑lite word‑level tokenization with WordPiece subwords and whole‑word masking (24 layers, 1024‑dim hidden, 16 heads, 32k vocab), with pretraining code on cl‑tohoku/bert‑japanese.
- bert-base-japanese-v3tohoku-nlp66kJapanese BERT‑base (12 layers, 768‑dim hidden, 12 heads, 32 k vocab) pretrained with whole‑word masking on CC‑100 and 2023‑Jan Wikipedia, using Unidic 2.1.2 word‑level tokenization plus WordPiece, in 2 M training steps.
- t5-base-japanese-v1.1sonoisa17kA Japanese T5‑v1.1 model pretrained on ≈100 GB of Wikipedia and OSCAR CC‑100 data (mixed 10:1 for SentencePiece with byte‑fallback), requiring fine‑tuning for downstream tasks, includes transfer‑learning sample code, notes potential bias in outputs, and is licensed CC‑BY‑SA 4.0.
- deberta-v3-base-japaneseku-nlp13kJapanese DeBERTa V3 base pre‑trained on 540 B tokens from LLM‑jp v1.0, trained with a modified DeBERTa V3 setup, uses a unigram byte‑fallback tokenizer (no morphological analyzer), and is fine‑tuned for JGLUE NLU tasks.
- t5-small-shortretrieva-jp3kA T5 v1.1 Japanese‑language model pre‑trained on mC4/ja and Wikipedia, featuring GEGLU activations, no dropout during pre‑training, separate embedding/classifier layers, and larger d_model with fewer heads, licensed CC‑BY‑SA 4.0 for commercial use.
Resources in similar research fields
- bert-base-japanese-whole-word-maskingtohoku-nlp462kJapanese BERT‑base pretrained on 2019 Japanese Wikipedia with IPA‑dictionary and whole‑word masking, 12‑layer 768‑dim, 32,000‑vocab, 512‑token sequences, 1 M steps, available at cl‑tohoku/bert‑japanese under CC‑BY‑SA.Shared: Pretrained encoders (BERT etc.)
- deberta-v2-large-japanese-char-wwmku-nlp76kJapanese DeBERTa V2 large model trained on 171 GB of Japanese Wikipedia, CC‑100, and OSCAR with character‑level sentencepiece tokenization and whole‑word masking, ready for downstream fine‑tuning through Hugging Face Transformers.Shared: Pretrained encoders (BERT etc.)
- bert-base-japanese-char-v2tohoku-nlp55kA BERT‑base Japanese model (12 layers, 768‑dim hidden states, 12 heads) trained on 30 M Wikipedia sentences (~4 GB) with Unidic 2.1.2 word‑level tokenization followed by character‑level tokenization and whole‑word masking, using 512‑token sequences, 256 batches, and 1 M training steps.Shared: Pretrained encoders (BERT etc.)
- modernbert-ja-310msbintuitions36kModernBERT‑Ja‑310M is a Japanese‑language BERT variant that blends local‑global attention with RoPE, trained on 4.09 T tokens of Japanese/English text, supports a 102 400‑word vocabulary, 8 192‑token sequences, and is optimized for Flash Attention 2.Shared: Pretrained encoders (BERT etc.)
- deberta-v2-base-japanese-char-wwmku-nlp24kA Japanese DeBERTa‑V2 base model pre‑trained on 171 GB of Japanese Wikipedia, CC‑100, and OSCAR text with character‑level tokenization, whole‑word masking, trained for 20 days on 8 A100 GPUs, and ready for downstream fine‑tuning.Shared: Pretrained encoders (BERT etc.)
- deberta-v2-base-japaneseku-nlp24kJapanese DeBERTa V2 base model pretrained on 171 GB of Japanese Wikipedia, CC‑100 and OSCAR data using Juman++ segmentation and SentencePiece tokenization, trained for three weeks on eight NVIDIA A100 GPUs and ready for fine‑tuning.Shared: Pretrained encoders (BERT etc.)