Nemotron-SFT-Science-v2: Synthetic and Non-Synthetic Science Problems with LLM Solutions

Nemotron-SFT-Science-v2 is a science reasoning dataset created by NVIDIA and last updated on June 4, 2026. It contains problems and solutions across three domains: Physics, Biology, and Chemistry. The dataset includes synthetic and vendor-sourced problems in multiple-choice and open-question formats, paired with LLM-generated solutions using chain-of-thought, Python, and search tool reasoning.

Use Cases

• Fine-tuning LLMs for science question answering based on the multiple-choice and open-question formats.
• Training models for chain-of-thought reasoning based on the CoT solution generation setup.
• Developing models that use external tools for problem-solving based on the Python and Tavily API tool usage setups.
• Benchmarking model performance across scientific domains based on the inclusion of Physics, Biology, and Chemistry problems.

Strengths

• Covers three distinct scientific domains: Physics, Biology, and Chemistry.
• Includes two question formats: multiple-choice questions (MCQ) and open questions (OpenQ).
• Provides solutions generated via three distinct reasoning setups: chain-of-thought, Python tool usage, and search tool usage.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset includes synthetic data, which may not fully reflect the complexity of real-world science problems.

Provenance

Source

NVIDIA

Collection Method

Combines synthetic generation (MCQ, RQA) and non-synthetic vendor problems with LLM-generated solutions.

Time Range

null

Freshness

Last updated 2026-06-04 04:50:29; freshness should be verified.

Geography

null