Coding datasets
Repository tasks for training and evaluating coding agents: issues, patches, and work in the terminal.
Request sampleOff-the-Shelf Frontier Datasets
Off-the-shelf datasets for training, evaluation, and RLVR – expert-built, quality-controlled, and ready for your stack.
-
Computer-use trajectoriesSFT · Eval
Screen recordings of real desktop jobs, split into actions with screenshots and generated reasoning.
View sample -
Grounded document Q&ASFT · Eval
Enterprise documents with fields and checked answers, from our work with Databricks on OfficeQA.
View sample -
Planning environmentsRLVR
Five RL environments for multi-step planning, scored with partial rewards.
Request sample
01 Computer use
Computer-use trajectories
Screen recordings of economically valuable work in desktop applications. Each recording captures one complete job, then is split into individual actions, each with its screenshot and generated reasoning.
The most wanted dataset by labs
- Task
- An economically valuable job in a desktop application, split into ordered subtasks.
- Recording
- The full job as one continuous screen recording.
- Steps
- The recording split into actions, each with its screenshot and the code that performed it.
- Reasoning
- A generated thought and observation for every step.
- Use
- Supervised fine-tuning and evaluation of computer-use agents.
02 Documents
Grounded document Q&A
Real enterprise documents paired with an overview, extracted fields, and questions whose answers are checked against the page.
From our work with Databricks on OfficeQA
- Page
- The source document, as scanned.
- Overview
- A short summary of what the document is.
- Fields
- Key-value pairs taken from the page.
- Questions
- Questions about the document, each answer checked against the page.
Databricks finalized the published OfficeQA benchmark separately, so this sample may differ from it.
More datasets
These collections are shared on request. Tell us what you are training, and we will send a sample.
Speech-to-speech
Paired spoken turns for models that listen and answer in speech.
Request sampleAudio transcription
Recorded speech with the transcript, for training and evaluating speech recognition.
Request sampleVideo
Video for models that follow an action across frames.
Request sampleRobotics
Demonstrations, sensor streams, failures, and recovery episodes for manipulation policies.
Request samplePlanning environments
RL environments for multi-step planning: employee scheduling, flights, private shuttles, amusement parks, and events. Scored with partial rewards.
Request sampleSpecify the dataset yourself.
Set the tasks, the environment, and who owns the result.
Frequently asked questions
Something else? Talk to our team
What's in a sample dataset?
A hand-picked set of real items from the full dataset, opened on the SuperAnnotate platform with every annotation, action, and field in place.
Do I need to talk to sales to view a sample?
No. Computer-use trajectories and grounded document Q&A open on the platform without a call. Planning environments and the other collections are shared on request.
Who creates the data?
People who do the work. Computer-use trajectories are screen recordings of people completing real jobs; the reasoning on each step is generated. Document questions are written and checked against the source page, the same kind of work we did with Databricks on OfficeQA. Planning environments and robotics demonstrations are built and reviewed by our team.
Can a dataset be extended for our use case?
Yes. Take a collection as published, or specify the tasks, the environment, and the domains. An extension can join the catalog at a discount, or stay with you.
Who owns a custom dataset?
You do. A custom dataset stays with you. Contributing the domain to the catalog is a discounted engagement, and that domain can then be licensed onward.
How is data delivered?
A sample opens on the SuperAnnotate platform. The licensed collection is the full dataset, delivered with the annotations, actions, and metadata on each item.