NVIDIA Nemotron Post-Training Dataset v2 and Nemotron Nano 2 9B Release

NVIDIA has released the Nemotron Post-Training Dataset v2, a collection of 6 million multilingual reasoning examples, alongside the Nemotron Nano 2 9B model. This release aims to support the open-weight model ecosystem by providing high-quality training data and an efficient edge-capable model with a configurable thinking budget to balance accuracy and cost.

Nemotron Nano 2 9B Model Specifications

The Nemotron Nano 2 9B is designed for edge and RTX deployments, specifically targeting analytics copilots, support chatbots, and customer service agents.

Technical Architecture and Performance

  • Hybrid Architecture: The model uses a hybrid Transformer–Mamba architecture (combining Mamba-2 with a limited number of attention layers), which allows for higher throughput compared to Transformer-only models of similar size.
  • Throughput: The model achieves up to 6x higher token generation throughput than other leading models in its 9B parameter size class.
  • Cost Optimization: A configurable "thinking budget" allows users to control the number of reasoning tokens used, which can reduce reasoning costs by up to 60%.
  • Licensing: The model is released under the nvidia-open-model-license.

Nemotron Post-Training Dataset v2 Construction

The Nemotron Post-Training Dataset v2 consists of 6 million reasoning examples translated from the original English reasoning data into five target languages: French, Spanish, German, Italian, and Japanese.

Translation Methodology

To leverage the English knowledge acquired during pre-training, NVIDIA translated the user prompts and model responses while preserving the original English reasoning chains. The translation process employed Qwen2.5-32B-Instruct-AWQ for German and Qwen2.5-14B-Instruct for other languages, selected for their robust quality, open Apache 2.0 license, and ability to fit on a single A100 GPU.

Quality Control and Hallucination Mitigation

NVIDIA identified that LLMs are more prone to hallucinations when translating SFT (Supervised Fine-Tuning) datasets compared to standard machine translation sets, and that quality degrades as input length increases. To mitigate this, the following mechanisms were implemented:

  • Line-by-Line Translation: Sentences are broken down by newline; non-translatable lines (such as tabs) and code blocks are preserved without translation.
  • Format Enforcement: Translations are required to be wrapped in specific brackets (– –). Examples that do not follow this format are discarded.
  • Language Identification: The fastText language ID tool is used on prompt inputs to filter out off-target data. This process resulted in the removal of 55,567 examples (approximately 1.1% of all multilingual examples).

Data Discard Rates by Format Enforcement

The following table details the ratio of data discarded (measured by bytes) due to the enforcement of the output format:

Language Code QA Math
German (de) 2.28% 1.11% 2.47%
Spanish (es) 26.14% 5.15% 6.38%
French (fr) 11.01% 1.37% 1.96%
Italian (it) 4.94% 1.36% 0.75%
Japanese (ja) 7.68% 2.51% 3.86%

Availability

The Nemotron Post-Training Dataset v2 is available on Hugging Face. The Nemotron Nano 2 9B model weights are also hosted on Hugging Face, with endpoints available via build.nvidia.com and as NVIDIA NIM for low-latency, high-throughput requirements.

Sources