NVIDIA has announced the release of a 6‑million‑item multilingual reasoning dataset, expanding on its earlier Nemotron Post‑Training Dataset V1. The new dataset translates the previously released English reasoning data into five target languages—French, Spanish, German, Italian, and Japanese—while preserving the original English reasoning chain in the user prompt and model response. The dataset is part of NVIDIA’s continued commitment to openness and transparency in model development, offering training data, tools, and final model weights to support the improvement of open‑weight models. In conjunction with the dataset, NVIDIA introduced the Nemotron Nano 2 9B, a hybrid Transformer–Mamba architecture designed for edge deployment. The Nano 2 9B allows users to configure a thinking budget, enabling adjustments to accuracy, throughput, and cost to meet real‑world needs. For the translation process, NVIDIA selected Qwen2.5‑32B‑Instruct‑AWQ for German and Qwen2.5‑14B‑Instruct for the other languages, based on benchmarking results. The release underscores NVIDIA’s focus on providing high‑quality, multilingual data to advance the capabilities of large language models and support the broader AI community.
NVIDIA Releases 6 Million Multi-Lingual Reasoning Dataset
NVIDIA has released a 6‑million‑item multilingual reasoning dataset, translating its earlier English reasoning data into French, Spanish, German, Italian, and Japanese. The dataset accompanies the Nemotron Post‑Training Dataset V2 and is intended to support open‑weight model development. NVIDIA also announced the Nemotron Nano 2 9B model, a hybrid Transformer–Mamba architecture designed for edge deployment with configurable accuracy, throughput, and cost.
Why it matters
Providing large, multilingual reasoning data enables broader training of open‑weight models, fostering innovation and reducing language bias in AI systems.