Wuxia Chinese-English translator
A MarianMT model fine-tuned on a 100k-pair wuxia/xianxia corpus (Bachelor's thesis, UDC), shown next to the base model it started from. Everything runs in your browser: nothing you type leaves your device. The first translation downloads each model (about 113 MB each, cached afterwards).
About the models
This page is a small test demo, not the best system of the thesis. It runs the lightest model (MarianMT, 77M parameters) because it is the only one that fits comfortably in a browser; the larger models need a GPU. The thesis compares eight fine-tuned models on the same wuxia test set, listed below.
Both models here are sentence-level: the text is split at punctuation and each sentence is translated on its own, without cross-sentence context. Sentence-final periods are added by this page. The browser versions are int8-quantized ONNX exports; the fine-tuned one was checked against the PyTorch output on sample sentences. The examples come from the thesis corpus.
| Model (fine-tuned) | Type | Parameters | SacreBLEU | chrF | COMET |
|---|---|---|---|---|---|
| MarianMT (this demo) | NMT, full fine-tuning | 77M | 5.1 → 28.4 | 23.6 → 50.0 | 59.1 → 77.6 |
| mT5-small | NMT, full fine-tuning | 300M | 0.0 → 17.0 | 1.3 → 39.4 | 0.0 → 68.9 |
| M2M100-418M | NMT, full fine-tuning | 418M | 5.1 → 33.1 | 25.2 → 53.2 | 56.2 → 78.5 |
| mBART-large-50 | NMT, full fine-tuning | 610M | 6.3 → 37.7 | 26.7 → 56.2 | 60.9 → 80.0 |
| Gemma 3 27B | LLM, QLoRA | 27B | 10.6 → 26.3 | 36.5 → 48.5 | 71.5 → 78.6 |
| Qwen3 30B-A3B | LLM (MoE), QLoRA | 30B (3B active) | 9.3 → 25.1 | 30.1 → 47.2 | 61.9 → 78.4 |
| GLM-4 32B | LLM, QLoRA | 32B | 9.6 → 13.5 | 36.3 → 39.9 | 72.5 → 72.4 |
| Llama 3.3 70B | LLM, QLoRA | 70B | 9.5 → 26.1 | 30.2 → 48.4 | 62.5 → 77.9 |
Each cell shows the score of the original base model, an arrow, and the score after fine-tuning (PyTorch models, wuxia test set, values from the model cards). Sorted by number of parameters.
Models: fine-tuned (PyTorch), fine-tuned (ONNX, this demo), base (ONNX, Xenova). Powered by Transformers.js.