nllb-200-1.3B-yi-en-ONNX

NLLB-200-1.3B, Yiddish ↔ English — ONNX

A Yiddish–English translation model derived from NLLB-200-distilled-1.3B, adapted for this single language pair and quantised to run in the browser with transformers.js in a 778 MB download. A larger sibling of the 600M build; the smaller model is what runs in production.

Quality, honestly

This is a draft translator, not a translation. On a held-out benchmark against a human-reviewed translation of oral-history speech it is substantially more faithful than the base model, but its characteristic failure is the dangerous one: fluent, confident English that gets a concrete noun wrong. In our own evaluation the base model rendered דאָרף (village) as "need"; models of this class do that.

Keep the Yiddish beside the English, and treat the English as a rough reading, not as the record. For material that matters — a family testimony someone will read as history — the English still needs a person.

Fine-tuning used machine-generated translations alongside human-translated sentence pairs. Human pairs from Tatoeba (CC-BY 2.0 FR).

Licence and attribution

NLLB-200 is the work of Meta AI, released under CC-BY-NC-4.0. This derivative carries the same licence and is therefore non-commercial. All credit for the underlying translation capability belongs to the NLLB team.

Methodology

The adaptation, training and evaluation methodology will be described in a forthcoming paper; this card will be updated with a citation when it is available.