---
title: Nvidia Nemotron V3 Data Atlas
emoji: 🗺️
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
short_description: Interactive embedding atlas for Nemotron post-training data
---


# Nemotron Post-Training v3 Prompt Atlas

An interactive atlas for exploring prompts from the **Nemotron-Post-Training-v3** dataset collection.

This Space provides a visual way to inspect the structure, diversity, and coverage of prompts used across post-training data. It is intended to help researchers, builders, and dataset curators quickly understand what kinds of tasks are represented in high-volume Nemotron open data for post-training.The atlas is generated from a volume-sampled set of prompt samples, preserving the relative ratio across datasets, then organized through semantic embeddings into an interactive visual map.

## What this Space helps with

Large post-training datasets are difficult to understand from tables alone. A prompt atlas makes the dataset easier to navigate by grouping semantically similar examples into a visual map.

Use this Space to:

* Browse prompt clusters across instruction following, reasoning, tool use, coding, math, safety, formatting, and other post-training domains  
* Inspect representative prompts from different regions of the dataset  
* Identify dense clusters, sparse regions and build intuition about the composition and diversity of Nemotron post-training data

## How to use it

### Explore the atlas

Pan or zoom around the map. Click or hover over a point to inspect its prompt and metadata.

### Refine the view

| Control | Location | Purpose |
|---|---|---|
| **Color overlays** | Top-left corner | Color points by `dataset`, `pipeline_stage`, or `domain`. |
| **Sidebar filters** | Right sidebar | Filter using Collections, Used in, Dataset, Pipeline stage, or Domain. |
| **SQL predicates** | Right sidebar | Create custom filters over the available metadata. |
| **Semantic search** | Lower-right corner | Find keywords and semantically similar queries. |

<p align="center">
  <img src="./gifs/explore_cropped.gif" alt="Explore the prompt atlas" width="42%" />
  <img src="./gifs/semantic_cropped.gif" alt="Search the prompt atlas" width="42%" />
</p>

<p align="center"><em>Explore the data using color overlays, filters, and semantic search.</em></p>

### Interpreting the atlas

Each point represents a prompt sample from a volume-sampled subset of the Nemotron post-training data. Volume sampling preserves the relative proportions of the source datasets in the broader data mixture, so the atlas provides a proportional view of the prompt space rather than an equal-sized sample from every dataset.

Prompt samples are embedded with [**nemotron-llama-embed-v2**](https://huggingface.co/nvidia/llama-nemotron-embed-1b-v2) and projected into an interactive visual map. The embedding model was used as a general-purpose semantic embedding model to support visualization. It provides one semantic lens for organizing prompt samples, rather than a canonical taxonomy of the data.

Nearby points generally indicate prompts that are semantically similar under this embedding model and projection pipeline. Because embedding choice can affect local neighborhoods and cluster boundaries, the atlas should be interpreted as a practical exploration view produced by this methodology.

The resulting layout is based on semantic similarity between individual prompt samples, not on dataset boundaries. As a result, a single dataset may appear across multiple regions of the map if it contains diverse task types, and a single visual cluster may contain examples from multiple datasets if their prompts are semantically similar.

The metadata overlays for `dataset`, `pipeline_stage`, and `domain` are intended to make these relationships easier to inspect directly.

The map is designed for qualitative exploration and dataset discovery. It is especially useful for identifying broad prompt families, comparing metadata overlays, inspecting representative examples, and finding regions that may warrant deeper review.

A few interpretation notes:

* Clusters reflect similarity in embedded prompt content, not necessarily dataset boundaries.  
* Dataset ratios are preserved through volume sampling, but individual datasets may appear in multiple semantic regions.  
* Distance is most useful for local exploration and cluster-level intuition, rather than as an exact numerical measure.  
* The atlas complements dataset cards, metadata, and model evaluations.

## Use this for Dataset Discovery

Quickly answer questions such as:

* What kinds of prompts are included?  
* Which task families dominate the dataset?  
* Are there visible clusters for math, code, tool use, safety, formatting, or agentic behavior?

The atlas is intended as a navigation and sensemaking layer over that data. For licensing, usage restrictions, provenance, and dataset-specific details, refer to the individual dataset cards in the collection.

## Citation and attribution

If you use this Space or the underlying datasets in your work, please cite the relevant NVIDIA Nemotron dataset cards, model cards, technical reports, or collection pages associated with the data.

## Feedback

Feedback, issues, and suggestions are welcome through the Hugging Face Space discussion tab.