Nemotron-Math-v2 Mathematical Reasoning Trajectories

## Overview Nemotron-Math-v2 is a large-scale mathematical reasoning dataset released by NVIDIA, containing approximately 347,000 high-quality mathematical problems paired with roughly 7 million model-generated reasoning trajectories. It accompanies the paper "Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode Supervision" (arXiv:2512.15489) and is distributed in Parquet format. The dataset supports long-context reasoning, tool use, and multi-mode supervision for training and distilling math-capable LLMs. ## Schema - problem — string — the mathematical problem statement - solution / trajectory — string — model-generated reasoning chain (often long-context) - answer — string — final answer (where applicable) - mode — string — supervision mode (e.g., CoT, tool-use) - source — string — origin or problem family - metadata — struct — additional per-record annotations - +additional columns documented on the source dataset card ## Sources - nvidia/Nemotron-Math-v2 on HuggingFace — https://huggingface.co/datasets/nvidia/Nemotron-Math-v2 — licensed CC-BY-4.0 and CC-BY-SA-4.0 - Companion paper: arXiv:2512.15489 - Code: NeMo-Skills (NVIDIA) ## Methodology NVIDIA curated ~347K mathematical problems from a range of public math sources and generated multiple reasoning trajectories per problem using strong teacher models under multi-mode supervision (e.g., chain-of-thought and tool-augmented modes). Trajectories are intended to support long-context distillation into smaller student models. Refer to the accompanying paper and NeMo-Skills documentation for full collection, filtering, and quality-control procedures. ## Known gaps & limitations The dataset is English-only and focused on mathematical reasoning, so it does not generalize to other domains or languages. Trajectories are synthetic (model-generated) and may contain reasoning errors, hallucinated steps, or biases inherited from the teacher model. Problem coverage overlaps with public math corpora (e.g., MATH, GSM8K-style data), creating potential leakage risk if used to train models evaluated on standard math benchmarks. Source does not exhaustively document deduplication against common eval suites; buyers should validate empirically before using for benchmark-adjacent training. ## Intended use & out-of-scope - Intended: fine-tuning and distilling mathematical reasoning into LLMs, long-context reasoning research, tool-use training, supervised trajectory learning, agent math evals. - Out-of-scope: not deduplicated against common math eval suites (MATH, GSM8K, AIME) — leakage risk for benchmark training; not suitable for non-math reasoning, non-English math, or as ground-truth answer key without verification. _Federated dataset: 5 parquet shards, 46.87 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: Nemotron-Math-v2: Mathematical Reasoning Trajectories (NVIDIA) NVIDIA's 347K math problems with 7M model-generated reasoning trajectories for distilling mathematical reasoning. Long-context, tool-use, multi-mode supervision. CC-BY-4.0/CC-BY-SA-4.0.