# DataPraxis releases verified analytics task dataset and models

> Researchers unveil DataPraxis-5K, a 5,000‑task dataset, and two new models that score 55.11 on DABStep and 51.9 on JobBench.

Oossa · 2026-10-21 · https://oossa.com/en/datapraxis-releases-verified-analytics-task-dataset-and-models

A team led by Nan Jia and Haojin Yang introduced DataPraxis, a framework that builds analytics tasks by verifying each step of a workspace. Using the framework they created DataPraxis‑5K, a collection of 5,000 verified analytical trajectories across many domains. They also trained two models, DataPraxis‑27B and DataPraxis‑35B‑A3B, which achieve 55.11 on the DABStep benchmark and 51.9 on JobBench, matching or beating comparable frontier models.

## The facts

- DataPraxis‑5K contains 5,000 verified task trajectories.
- DataPraxis‑27B scores 55.11 on DABStep and 51.9 on JobBench.

## Why it matters

The dataset gives researchers reliable training material for building AI that can handle real‑world analytics work.

## Sources & references

1. [DataPraxis: Scaling Analytics Tasks via Verified Workspace Evolution](https://arxiv.org/abs/2610.26924) – arXiv, 2026-10-21

Last updated: 2026-10-21