MameLoshnLM: First Open-Source 8B LLM and Benchmark for Yiddish
Yiddish-NLP · hf · 2026-08-07
The Yiddish-NLP team released MameLoshnLM, the first open-source 8B-parameter language model built specifically for Yiddish. Addressing the scarcity of reliable digital resources and the prevalence of noisy, machine-translated text in existing datasets, the team also introduced Oytser, a high-quality pretraining corpus, and Kashes, a multi-task benchmark.
Using these resources, the researchers continued pretraining Llama 3.1 8B. MameLoshnLM outperformed open baselines of similar scale across tasks in the Kashes benchmark, including translation, linguistic analysis, and information extraction.
Beyond quantitative gains, the model better captures language-defining lexical and morphological patterns. The research highlights a broader failure mode of noisy web-scale multilingual data for low-resource languages, providing a practical template for developing models for historically rich but digitally underrepresented languages.
More from Models
- Qwen 3.8-Max Set to Drop Next Week: Starting with 2.4T Parameters, 27B to Follow — Ok-Shower7286 · 2026-08-07
- SK Telecom Releases A.X K2: A 688B Parameter Open-Source MoE Model — huggingface · 2026-08-07
- NVIDIA Exec: Reasoning Model Alpamayo Solves Self-Driving's Long Tail Problem — ZGojcic · 2026-08-07
- DeepSeek-V4-Flash Broken on AMD MI325X? User Reports Tool Call Chaos — Brunofcsampaio · 2026-08-07
- ChatGPT Criticized for 'Pseudo-Empathy': Users Slam Toxic Neutrality in Emotional Advice — Purelig · 2026-08-07
- WeirdML v2 Benchmark Released: Tasks Expanded to 19, Clear Cost-Performance Scaling — yacineMTB · 2026-08-07