MameLoshnLM: Yiddish Language Model and Evaluation Benchmark
The paper introduces the first open-source 8B Yiddish-specific language model, plus new training and evaluation resources.
MameLoshnLM is built by continuing pretraining Llama 3.1 8B on Oytser, a curated Yiddish corpus mixing web-native and literary material. The authors also release Kashes, a benchmark covering translation, linguistic analysis, information extraction, and language understanding. They report that MameLoshnLM beats open baselines of similar size and better captures Yiddish lexical and morphological patterns than general multilingual models. HF Daily Papers' note
MameLoshnLM is built by continuing pretraining Llama 3.1 8B on Oytser, a curated Yiddish corpus mixing web-native and literary material. The authors also release Kashes, a benchmark covering translation, linguistic analysis, information extraction, and language understanding. They report that MameLoshnLM beats open baselines of similar size and better captures Yiddish lexical and morphological patterns than general multilingual models. HF Daily Papers' note
score 5