---
pretty_name: Nemotron Math V2 Low Mini
task_categories:
- text-generation
language:
- en
license:
- cc-by-4.0
- cc-by-sa-4.0
size_categories:
- 10K<n<100K
dataset_info:
  features:
  - name: id
    dtype: string
  - name: source
    dtype: string
  - name: question
    dtype: string
  - name: steps
    list: string
  - name: answer
    dtype: string
  splits:
  - name: train
    num_bytes: 136634937
    num_examples: 58595
  download_size: 99758066
  dataset_size: 136634937
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# Nemotron Math V2 Low Mini

A compact textual reasoning dataset derived from the `low` split of [`nvidia/Nemotron-Math-v2`](https://huggingface.co/datasets/nvidia/Nemotron-Math-v2) at immutable revision `8e793210e175b6406c752a870f585f62de98c0d3`.

## Selection boundary

This extraction intentionally excludes tool-use semantics:

- Accept exactly two messages with roles `user` then `assistant`.
- Reject records declaring tools, containing assistant tool calls, containing tool-result messages, or containing any additional turns.
- Require the injected source user message to end with the exact `problem` text.
- Read reasoning only from the sole assistant message's `reasoning_content`.
- Read the final response only from that assistant message's `content`.
- Preserve pinned `low`-split row order without sampling or shuffling.

Each output row contains `id`, `source`, `question`, `steps`, and `answer`. Reasoning is split on blank lines into whitespace-stripped steps. Source UUIDs are not globally unique across trajectories, so every output ID combines the immutable zero-padded source row position with the source UUID.

Rows must contain 3-50 steps, at least 128 total reasoning characters, and no step longer than 1,024 characters. The exact Llama 3 training sequence must be at most 4,096 characters and 1,024 tokens under the prepared `models/llama3-1.2b` tokenizer.

## Dataset statistics

| Metric | Value |
|---|---:|
| Source rows scanned | 1,718,159 |
| Final records | 58,595 |
| Retention | 3.410% |
| File size (canonical JSONL) | 145,436,648 bytes |
| Sources represented | 2 |
| Total reasoning steps | 328,464 |
| Mean steps per record | 5.61 |
| Median steps per record | 5 |
| Steps per record | 3-46 |
| Mean individual step length | 179.68 characters |
| Median individual step length | 103 characters |
| Individual step-length range | 2-1,024 characters |
| Median total reasoning length | 902 characters |
| Total reasoning-length range | 129-3,698 characters |
| Answer-length range | 9-3,143 characters |
| Question-length range | 14-1,102 characters |

## Source composition

| Source | Records |
|---|---:|
| `nvidia/Nemotron-Math-v2/stackflow` | 38,857 |
| `nvidia/Nemotron-Math-v2/aops` | 19,738 |
| **Total** | **58,595** |

## Rejection accounting

| Reason | Rows |
|---|---:|
| Source conversation is not exactly two turns | 462,352 |
| Tool trajectory | 412,640 |
| Fewer than 3 reasoning steps | 352,792 |
| Complete Llama 3 sequence over 4,096 characters | 180,422 |
| Complete Llama 3 sequence over 1,024 tokens | 134,549 |
| A reasoning step over 1,024 characters | 106,411 |
| More than 50 reasoning steps | 10,290 |
| User message does not end with exact problem | 103 |
| Fewer than 128 reasoning characters | 5 |
| **Accepted** | **58,595** |
| **Total** | **1,718,159** |

## Integrity and provenance

The canonical local JSONL has SHA-256 `5bc93ca130c45f1b94df993cc5f55127d2431b4010655ec6b2ee40dc4a39189a`. The repository includes `construction_manifest.json` with the pinned source revision, tool-exclusion boundary, ID scheme, complete filter contract, rejection counts, and content hash.

Rows retain the source dataset's row-level Creative Commons licensing. Usage remains subject to the source dataset's licensing terms and documented limitations.