This repository contains an English-to-Chinese MarianMT model fine-tuned from Helsinki-NLP/opus-mt-en-zh, plus the checkpoint of a smaller Transformer trained from scratch for comparison.

Intended use

Educational English-to-Chinese sentence translation
Reproducing the accompanying Transformer/MarianMT comparison
Local inference and further research fine-tuning

This is not a production translation service. It was trained on short conversational sentences and can mistranslate long, technical, safety-critical, or culturally sensitive text.

Evaluation

The fixed course test split contained 2,991 sentence pairs. Scores were computed with SacreBLEU (tokenize="zh"), chrF++ and TER.

Model	BLEU	chrF++	TER	Exact match
Scratch Transformer	20.5691	16.0240	66.1049	3.0090%
Base MarianMT (zero-shot)	41.7833	28.7659	41.7261	2.1398%
Fine-tuned MarianMT	52.4721	39.2654	30.8395	17.8536%

These scores describe this fixed local split only and should not be interpreted as general-domain translation quality.

Training details

Base model: Helsinki-NLP/opus-mt-en-zh
Original course corpus: 26,918 training pairs, split into 25,573 train and 1,345 validation pairs
Epochs: 4; selected checkpoint: epoch 3 by validation loss
Optimizer: AdamW
Learning rate: 2e-5 with cosine decay
Micro-batch: 16; gradient accumulation: 2
Beam size: 4
Seed: 2026
Training device: NVIDIA GeForce RTX 4060 Laptop GPU

The sentence content appears to derive from the Tatoeba/ManyThings English–Mandarin collection. The course files did not retain per-sentence attribution metadata, so the corpus itself is not redistributed here. See the project repository for the expected data format and reproducibility notes.

Usage

python

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "J1mmymm/transformer-en-zh-translator"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

inputs = tokenizer("Machine learning is useful.", return_tensors="pt")
outputs = model.generate(**inputs, num_beams=4, max_new_tokens=64)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Additional scratch checkpoint

scratch/scratch_transformer.pt contains the state dictionary, English and Chinese vocabularies, exact architecture/training configuration, best validation loss, and best epoch for the custom nn.Transformer model. Load it with en_zh_translator.load_scratch_checkpoint from the GitHub project.

Limitations and bias

The model inherits limitations and biases from the base model and the short conversational fine-tuning corpus. References may have multiple valid translations, so exact-match is a particularly strict and incomplete metric. Human review is required before consequential use.

Licensing and attribution

The base model is Apache-2.0. Project code and this derived model release are published under Apache-2.0. Dataset text is not included; users preparing their own Tatoeba/ManyThings split must follow its attribution requirements.