This repository provides Japanese ModernBERT trained by SB Intuitions.
ModernBERT is a new variant of the BERT model that combines local and global attention, allowing it to handle long sequences while maintaining high computational efficiency. It also incorporates modern architectural improvements, such as RoPE.
Our ModernBERT-Ja-130M is trained on a high-quality corpus of Japanese and English text comprising 4.39T tokens, featuring a vocabulary size of 102,400 and a sequence length of 8,192 tokens.
For all models, the vocabulary size is 102,400, the head dimension is 64, and the activation function is GELU. The configuration for global attention and sliding window attention consists of 1 layer + 2 layers (global–local–local). The sliding window attention window context size is 128, with global_rope_theta set to 160,000 and local_rope_theta set to 10,000.
Training with 450B tokens, comprising high-quality Japanese data.
The data consists of 150B tokens, and we trained it for 3 epochs.
This is because the overall performance of the Japanese language task improved with 3 epochs compared to just 1 epoch.
The sequence length is 8,192 without sequence packing.
Masking rate is 15% (with 80-10-10 rule).
The key differences from the original ModernBERT are:
It is pre-trained on Japanese and English corpora, leading to a total of approximately 4.39T training tokens.
We observed that decreasing the mask rate in Context Extension Phase 2 from 30% to 15% improved the model's performance.
Our model boasts a large vocabulary size of 102,400, which is larger than most existing Japanese models. To align the number of parameters with existing models, we set the hidden size to 512 and the number of hidden layers to 19. Finally, the model has 52M parameters in the embedding layer, 80M parameters in the Transformer layers, and a total of 132M parameters.
We use the tokenizer and vocabulary from sbintuitions/sarashina2-13b. Specifically, we employ a SentencePiece tokenizer with a unigram language model and byte fallback.
We do not apply pre-tokenization using a Japanese tokenizer. Therefore, users can directly input raw sentences into the tokenizer without any additional preprocessing.
You can use this model for masked language modeling, but it's mostly intended to be fine-tuned on a downstream task. Note that this model is not designed for text generation. When you want to generate a text, please use a text generation model such as Sarashina.
Since the unigram language model is used as a tokenizer, the token boundaries often do not align with the morpheme boundaries, resulting in poor performance in token classification tasks such as named entity recognition and span extraction.
These tasks are short-sequence evaluation tasks, and we aligned our settings with those of existing models. While the maximum sequence length varies across tasks, it does not exceed 512. We set the sequence length and other experimental configurations per task, ensuring that the settings remain consistent across models.
For hyperparameters, we explored the following ranges:
In the experiments, we loaded several Japanese models that are publicly available on HuggingFace using AutoModel and constructed classification models by appending a classification head consisting of a linear layer, a GELU activation function, and another linear layer. This was done because HuggingFace's AutoModelForSequenceClassification comes with different implementations for each model, and using them directly would result in classification heads that differ from one model to another.
For the embeddings fed into the classification layer, we used the embedding of the special token at the beginning of the sentence. That is, [CLS] in BERT and <s> in RoBERTa. Note that our model does not perform the next sentence prediction (NSP) task during pretraining, so <s> is added at the beginning of the sentence, not <cls>. Therefore, we used the <s> token for classification.
We conducted evaluations using 5-fold cross-validation. That is, we trained the model on the train set and evaluated it on the validation set. After determining the optimal hyperparameters (learning rate, epochs) based on the average performance on the validation sets, we report the average performance on the test sets with the hyperparameters.
For datasets without predefined splits, we first set aside 10% of the data as the test set and then performed 5-fold cross-validation on the remaining data. For datasets such as some tasks in JGLUE, where only train and validation sets are publicly available, we treated the validation set as the test set and performed 5-fold cross-validation on the remaining data. For datasets with predefined train, validation, and test sets, we simply trained and evaluated the model five times with different random seeds and used the model with the best average evaluation score on the validation set to measure the final score on the test set.
The evaluation results are shown in the table. #Param. represents the number of parameters in both the input embedding layer and the Transformer layers, while #Param. w/o Emb. indicates the number of parameters in the Transformer layers only.
Our ModernBERT-Ja-130M, a base-sized model, outperformed Tohoku BERT-large and achieved performance comparable to LUKE-japanese-large-lite. Specifically, it demonstrated impressive results in knowledge-based tasks such as JCommonsenseQA and RCQA.
Despite being a long-context model capable of processing sequences of up to 8,192 tokens, our ModernBERT-Ja-130M also exhibited strong performance in short-sequence evaluations.
ModernBERT-Ja-130M may produce representations that reflect biases. When you use this model for masked language modeling, it may generate biases or harmful expressions.
onnx-community