Modelpublic

mathew-felix/en-es-nmt-transformer

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes

Model Card

English-Spanish Custom Transformer Artifacts

This repository stores model artifacts for the english-spanish-translator Institutional Translation Review demo.

Important: this is a custom PyTorch checkpoint, not a standard Hugging Face Transformers checkpoint. It should not be loaded with AutoModel.from_pretrained. The serving code in the project repository loads the checkpoint through source.inference.InferenceEngine.

Artifacts

—best_model.pth: custom encoder-decoder Transformer checkpoint.
—tokenizer/: tokenizer files used by the project runtime.

Hosted Demo

The public Gradio Space is:

https://huggingface.co/spaces/mathew-felix/nmt-translator

The Space can run in full-model mode by downloading these artifacts, or in demo mode with deterministic sample outputs.

Intended Use

This model is one component of an institutional translation review workflow. The demo shows a custom model draft, optional bilingual evidence, review decision, fallback status, and final wording.

It is not presented as a replacement for DeepL, Google Cloud Translation, Azure Translator, GPT, MarianMT, or professional translation memory tools.

Loading

Use the project code rather than AutoModel:

bash

python scripts/download_model.py --source huggingface --hf-repo-id mathew-felix/en-es-nmt-transformer
python -m uvicorn src.serve:app --host 127.0.0.1 --port 8000

Limitations

—The checkpoint is about 1.27 GB.
—CPU inference can be slow.
—The model is weaker and slower than MarianMT on the included comparison run in the project README.
—Retrieval and GPT review are separate optional systems; they are not included in this model repository.