# sparkastML NMT

A set of models that aims to offer best open-source machine translation, based on the [OpenNMT](https://opennmt.net/).

## News

sparkastML's translation model is now updated!

### Details

- **Source Language:** Chinese (Simplified)
- **Target Language:** English
- **Training Time:** Totally 11.3 hours, 46,500 steps (~1×10¹⁸ FLOPs)
- **Training Device:**
  - RTX 3080 (20GB): 0-20,000 steps
  - RTX 4070: 20,000-46,500 steps
- **Corpus Size:** Over 10 million sentences
- **Validation BLEU Score:** 21.28
- **Validation Loss (Cross Entropy):** 3.152

### Model Download

Avaliable soon.

### Special thanks

[yumechi](https://github.com/eternal-flame-AD/) for sponsoring an RTX 4070 for training.

## History

### Sep 19, 2024

sparkastML's translation model is now updated!

#### Details

- **Source Language:** Chinese (Simplified)
- **Target Language:** English
- **Training Time:** 5 hours, 20,000 steps
- **Training Device:** RTX 3080 (20GB)
- **Corpus Size:** Over 10 million sentences
- **Validation BLEU Score:** 17
- **Version:** 1.0

#### Model Download

- **Google Drive:** [Download from Google Drive](https://drive.google.com/drive/folders/1-q_AKfQENW-pV6uAleUHPE9ghddfNWKF)
- **IPFS:** [Download from IPFS](http://ipfs.a2x.pub/ipfs/QmUMadzkBwvH5KTpoxfv7TgqzaPpqBzkXtkecV9TXPfZ3F/)
  - **CID:** `QmUMadzkBwvH5KTpoxfv7TgqzaPpqBzkXtkecV9TXPfZ3F`
- **GitHub Release:** [Go to Release Page](https://github.com/alikia2x/sparkastML/releases/tag/v2-model)