foundation

Training

Datasets that teach models to act.

Training is a data problem before it is a model problem. We build the datasets that show a model how to complete a task, not just how to answer a question.

What you get

Task Datasets

Collections built around real work: a goal, the steps that reach it, and the tool calls in between, kept in the order they actually happened.

Expert Trajectories

Recordings of people doing the job, like coding, reviewing, and fixing, with the reasoning and the mistakes left in rather than edited out.

Evaluation Datasets

Held-out tasks scored on whether the work got done, so a dataset can be checked and trusted before anyone trains on it.

Published datasets

Find our datasets on Hugging Face.

Browse all datasets

HTML Knowledge Agent

HTML_Knowledge-Agent

HTML and web-development knowledge for agentic coding workflows.

View dataset

Terminal Knowledge

TerminalKnowledge

Cross-platform terminal command references and agentic coding task data.

View dataset

AvBench 1.0

AvBench1.0_BENCHMARK

A benchmark dataset for evaluating aviation-focused model performance.

View dataset

Aviation Message Decoding

NOTAM-TAF-SPECI-METAR-ACARS_Decode

Aviation message and weather-format decoding data, including NOTAM, TAF, SPECI, METAR, and ACARS.

View dataset

Tower-1 ATC

Tower-1_ATC

Air traffic control-focused data for aviation language and ATC model workflows.

View dataset

AeroEmbed-1.5 Validation Checkpoint

aeroembed1.5_92726_validationcheckpoint

Published validation checkpoint data associated with AeroEmbed-1.5.

View dataset

AeroEmbed-1 Validation Checkpoint

aeroembed1_92626_validationcheckpoint

Published validation checkpoint data associated with AeroEmbed-1.

View dataset

Python Knowledge Agent

PYTHON_Knowledge-Agent

Python knowledge and programming data for coding-agent training.

View dataset

Science & Philosophy

SciencePhilosophy

Science and philosophy knowledge data.

View dataset

How it works

  1. 01

    Define the task

    We write it with you: what counts as done, and what a good run looks like.

  2. 02

    Collect and verify

    Trajectories are gathered and checked before they enter a set, so the signal stays clean.

  3. 03

    Train and measure

    Run the dataset through your model, then score it against the eval set before and after.

You can't improve what you haven't measured.

Talk to usSee the full stack