VanModers/oostfraeisk_ooversetter
First AI-based translator for East Frisian Low Saxon (https://oostfraeisk.org/translator), based on the Helsinki-NLP/opus-mt-de-nl model (https://huggingface.co/Helsinki-NLP/opus-mt-de-nl). Developed by the Oostfräisk Instituut: www.oostfraeisk-instituut.org
What's novel
First AI-based translator for East Frisian Low Saxon (https://oostfraeisk.org/translator), based on the Helsinki-NLP/opus-mt-de-nl model (https://huggingface.co/Helsinki-NLP/opus-mt-de-nl). Developed by the Oostfräisk Instituut: www.oostfraeisk-instituut.org
Code Analysis
14 files read · 4 roundsFine-tunes Meta's NLLB-200 multilingual model to support East Frisian Low Saxon (a minority language with no native NLLB token) by registering a new language embedding seeded from Dutch, generating ~150k parallel training pairs from a 111k-entry dictionary database, and providing a Gradio/CLI transl
Strengths
The add_frs_lang trick of exploiting spare embedding rows to add a new language without resizing is genuinely clever and well-documented. The data pipeline is thorough — multiple generation strategies (phrases, smashed pairs, noun templates, fan texts), Tatoeba cross-referencing for trilingual data, human-in-the-loop dictionary verification, and proper train/val splitting before bidirectional expansion to prevent leakage.
Weaknesses
No automated tests exist at all — the only 'testing' is a validator script that runs BLEU on a small validation set. The codebase has significant duplication (pretrained_de_frs and pretrained_frs_de are near-identical, the C# and Python dataset creators overlap heavily), and the trainer.py hardcodes a sample sentence for evaluation rather than using a proper eval harness.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
94
Contributors
1
Files
58
Active weeks
19
Repository
Language
Python
Stars
1
Forks
0
License
MPL-2.0