Open fine-tuning datasets for Japanese and Portuguese, alongside English

Pair translated reasoning data with locally grounded examples

Start with Nemotron-SFT-Multilingual-v2, an NVIDIA supervised fine-tuning (SFT) dataset that adds Brazilian Portuguese, Hindi, and Korean, plus refreshed Japanese data. It is built by translating NVIDIA math, code, and science datasets, so it carries reasoning skills into each language. Its card recommends that teams who also use Nemotron-SFT-Multilingual-v1 swap in the v2 Japanese data, merge the two releases, and balance the languages. For European Portuguese, the AMALIA technical report shows a working recipe. A consortium of Portuguese universities built AMALIA, a fully open 9B model for European Portuguese. NVIDIA open data makes up about 63% of its supervised fine-tuning tokens: Nemotron-Post-Training-Dataset v1 and v2 supply 57%, and three pt-PT datasets the team generated with Nemotron-Personas-USA supply about 6%.

Translated data does not show how your customers speak. Country-grounded personas fill that gap. Nemotron-Personas datasets publish synthetic personas aligned to real population distributions for individual countries, including Japan and Brazil. Teams use them as seeds to generate training examples with local names and everyday situations. NTT DATA, for example, used 500 personas from Nemotron-Personas-Japan to expand 450 seed examples into more than 138,000 training examples for its tsuzumi 2 model, and legal Q&A accuracy rose from 15.3% before training to 79.3% (NTT DATA case study). NVIDIA NeMo Data Designer, an open-source synthetic data tool, can generate these examples from a persona seed. Keep English data in the blend so English quality holds while Japanese and Portuguese improve, and test each language on its own evaluation set.

Datasets to start with

• Nemotron-SFT-Multilingual-v2: translated reasoning SFT data with Brazilian Portuguese, Hindi, Korean, and refreshed Japanese.

• Nemotron-Personas-Japan and Nemotron-Personas-Brazil: country-grounded synthetic personas for generating local examples.

• Nemotron-Personas collection: every published country edition in one place.

• HelpSteer3: human preference data that includes Japanese and Portuguese, for a preference-tuning pass after SFT.

Key takeaways

• Nemotron-SFT-Multilingual-v2 adds Brazilian Portuguese and refreshed Japanese SFT data.

• Translated reasoning data and locally grounded examples solve different problems. Use both.

• Nemotron-Personas country editions, including Japan and Brazil, seed synthetic examples with local context.

• NVIDIA open data makes up about 63% of the fine-tuning tokens behind AMALIA, a fully open European Portuguese model.

• Keep English data in the blend and test each language separately.