| Property | Value |
|---|---|
| Parameter Count | 122M |
| License | CC-BY-SA-4.0 |
| Training Data | Wikipedia, CC-100, OSCAR |
| Token Type | Character-level |
| Author | ku-nlp |
What is deberta-v2-base-japanese-char-wwm?
This is a specialized Japanese language model based on the DeBERTa V2 architecture, designed with character-level tokenization and whole word masking (WWM). Trained on a massive dataset of 171GB of Japanese text, it represents a significant advancement in Japanese natural language processing capabilities.
Implementation Details
The model was trained using 8 NVIDIA A100-SXM4-40GB GPUs over 20 days, implementing a sophisticated training procedure with sentencepiece tokenization (22,012 tokens). The training utilized a linear learning rate schedule with warmup, reaching completion at 320,000 steps.
- Learning rate: 2e-4 with Adam optimizer
- Batch size: 2,208 (total)
- Sequence length: 512 tokens
- Training corpus: 171GB combined data
Core Capabilities
- Masked Language Modeling
- Character-level tokenization
- Whole Word Masking
- Fine-tuning support for downstream tasks
Frequently Asked Questions
Q: What makes this model unique?
This model combines character-level tokenization with whole word masking, making it particularly effective for Japanese text processing. Its training on a diverse dataset including Wikipedia, CC-100, and OSCAR provides robust language understanding capabilities.
Q: What are the recommended use cases?
The model excels in masked language modeling tasks and can be fine-tuned for various downstream applications like text classification, named entity recognition, and question answering in Japanese language contexts.