OFF-THE-SHELF

Training & Validation
Off-the-Shelf Data

Datasets built with domain experts and tested on open-weight models, so you can see the lift before you commit.
Drop them into your pipeline and start iterating this week.

Get the Catalog

Training results

  • BioMysteryBench

    DeepSeek-V4-Flash, measured by pass@1

    Before training 73.30% , after training 81.10%

    Improvement +7.8pp

    Request the Full Catalog

  • JobBench

    DeepSeek-V4-Flash, measured by rubrics score

    Before training 48.98% , after training 62.03%

    Improvement +13.05pp

    Request the Full Catalog

  • Automation Bench

    Qwen3.8-27B, measured by pass@1

    Before training 39.20% , after training 49.20%

    Improvement +10pp

    Request the Full Catalog

  • Frontier Science-Research

    GLM-5.3, measured by success rate

    Before training 25.00% , after training 37.78%

    Improvement +12.78pp

    Request the Full Catalog

  • Apex-agent v1.0

    Qwen3.8-27B, measured by pass@1

    Before training 33.50% , after training 40.20%

    Improvement +6.7pp

    Request the Full Catalog

  • SciCode Verified

    Qwen3.8-27B, measured by Subtask avg@3 · With background

    Before training 82.93% , after training 92.10%

    Improvement +9.17pp

    Request the Full Catalog

Looking for more?

Our data catalog extends beyond what’s shown here. Contact us at business@unipat.ai.