Downloads · 30 days
14
2% of all-time downloads
wstcg-mt-ja-en
eepj/wstcg-mt-ja-en
wstcg-mt-ja-en is a translation model from eepj. Use it when you need text moved from one language to another. It is set up for transformers.
A Japanese-English machine translation model specifically trained for translating card text from the Weiss Schwarz (WS) Trading Card Game, fine-tuned on Helsinki-NLP/opus-mt-ja-en.
All-time downloads
844
Public
Parameters
75.3M
1.8 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors301 MB · 99%
At a glance
- Task
- Translation
- Library
- transformers
- Model type
- marian
- Access
- Public
- Created
- Apr 20, 2024
- Updated
- May 22, 2024
- SHA
- 431f0306
Try a prompt
Base models
Spaces
Model card
From the Hugging Face model README
WS TCG Card Text Translator
A Japanese-English machine translation model specifically trained for translating card text from the Weiss Schwarz (WS) Trading Card Game, fine-tuned on Helsinki-NLP/opus-mt-ja-en.
Gradio App
Hugging Face Spaces
Check out the Gradio app on Hugging Face Spaces https://huggingface.co/spaces/eepj/wstcg-mt.
[!Warning] Expect slow inference.
Running Locally
Clone the Hugging Face Spaces repository and run app.py.
git clone https://huggingface.co/spaces/eepj/wstcg-mt
cd wstcg-mt
python3 app.py
Dataset
Official WS Card List
- Japanese-English parallel card text comprising 6000+ card text retrieved from the offical card list.
Training
Base Model
- Base model: Helsinki-NLP/opus-mt-ja-en
- Base tokenizer: Helsinki-NLP/opus-mt-ja-en
- Source language: Japanese (ja)
- Target language: English (en)
Additional Tokens
| Token Type | Additional Tokens |
|---|---|
| Named Entity Placeholders | <TRAIT>, <NAME> |
| Trigger Icon Placeholders | <SOUL>, <CHOICE>, <TREASURE>, <SALVAGE>, <STANDBY>,<br> <GATE>, <BOUNCE>, <STOCK>, <SHOT>, <DRAW> |
| Keywords | 【, 】, AUTO, ACT, CONT, COUNTER, CLOCK, トリガー |
Hardware
- NVIDIA RTX3060 Ti with CUDA hardware acceleration
Hyperparameters
- Number of epochs: 5
- Optimizer: Adam
- Initial learning rate: 1e-4
- Learning rate scheduler: StepLR, reduce by factor of 0.5 every epoch
- Batch size: 4
- Loss function: CrossEntropyLoss
- Random seed: 42
Performance
Metrics
| Dataset | BLEU | chr-F |
|---|---|---|
| WS Official Card List | 0.82664 | 0.96515 |
Example Test Case
| Language | Official Text |
|---|---|
| Japanese | 【永】 あなたの<TRAIT>のキャラが4枚以上なら、このカードは、色条件を満たさずに手札からプレイでき、あなたの手札のこのカードのレベルを－1。 |
| English | 【CONT】 If you have 4 or more <TRAIT> characters, this card gets -1 level while in your hand, and can be played from your hand without fulfilling color requirements. |
| Model | Translated Text | BLEU | chr-F |
|---|---|---|---|
| Ours | 【CONT】If you have 4 or more <TRAIT> characters, this card can be played from your hand without fulfilling color requirements, and this card gets -1 level while in your hand. | 0.79134 | 0.95207 |
| Google Translate | [CONT] If you have 4 or more characters in <TRAIT>, you can play this card from your hand without meeting the color conditions, and the level of this card in your hand becomes -1. | 0.26561 | 0.53877 |
| GPT-3.5 | [Permanent] If you have 4 or more cards with the <TRAIT> trait, you can play this card from your hand without meeting the color condition, and reduce the level of this card in your hand by 1. | 0.24213 | 0.48925 |
| opus-mt-ja-en | If there are more than four characters in your <TRAIT> card, this card can be played from your hand with no color conditions, and you can see the level of this card in your hand is -1. | 0.22006 | 0.51034 |
References
Helsinki-NLP/opus-mt-ja-en <br> https://huggingface.co/Helsinki-NLP/opus-mt-ja-en
WS Official Card List (Japanese) <br> https://ws-tcg.com/cardlist
WS Official Card List (English) <br> https://en.ws-tcg.com/cardlist
WS Comprehensive Rules (English) <br> https://en.ws-tcg.com/wp/wp-content/uploads/20220726205726/WSE-Comprehensive-Rules-v2.07.pdf
- Task
- Translation
- Library
- transformers
- Type
- marian
- Languages
- ja, en
- Created
- Apr 20, 2024
- Updated
- May 22, 2024