Yiddish-NLP/MameLoshnLM

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 27, 2025License:cc-by-nc-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

MameLoshnLM is an 8 billion parameter language model developed by Yiddish-NLP, built by continued pretraining of Meta's Llama 3.1 8B architecture. It is specifically adapted for Yiddish, trained on a curated native-Yiddish corpus combining contemporary web sources and literary materials. This model excels in Yiddish linguistic tasks, understanding, named entity recognition, and machine translation, making it the first large generative language model tailored for the Yiddish language. It is intended for completion-style prompting, few-shot in-context learning, and as a base for further task-specific fine-tuning.

Loading preview...

MameLoshnLM: A Yiddish-Specific Language Model

MameLoshnLM is an 8 billion parameter language model developed by Yiddish-NLP, representing the first large generative model specifically adapted for Yiddish. It is built upon the Meta Llama 3.1 8B architecture, undergoing continued pretraining on a comprehensive Yiddish corpus that includes both contemporary web-native sources and extensive literary materials. The model maintains the original Llama 3.1 tokenizer and vocabulary.

Key Capabilities & Performance

This model demonstrates strong performance across a broad Yiddish benchmark suite, often outperforming its base Llama 3.1 8B and Qwen3 8B on Yiddish-specific tasks. Key strengths include:

  • Linguistic Tasks: Achieves high accuracy in POS tagging (88.6%), dependency parsing (40.6% LAS), lemmatization (31.9% change accuracy), and transliteration (92.3% 1-CER).
  • Named Entity Recognition (NER): Shows superior Micro F1 scores on EHRI (41.3%), WikiANN (59.7%), and newNLP (57.6%) datasets.
  • Machine Translation: Significantly outperforms other models in both English-to-Yiddish and Yiddish-to-English translation, with COMET scores of 78.5 and 87.2 respectively on FLORES+.

Usage & Limitations

MameLoshnLM is a base (non-instruct) model designed for completion-style prompting, few-shot in-context learning, and as a foundation for task-specific fine-tuning. It inherits limitations from its base model and training data, including potential reproduction of OCR artifacts, historical orthographic variations, and topical biases due to Yiddish being a low-resource language. It has not been instruction-tuned or safety-tuned. The model is released under a non-commercial CC BY-NC 4.0 license, with additional restrictions from the Llama 3.1 Community License.

Finetunes
1 models