Enterprise Atlas Commercial
RL training environments built with real enterprise data, where agents complete end-to-end tasks inside a realistic company. Designed to show where frontier models still fall short on business workflows.
Frontier Training Data. Ready Today.
Frontier-grade training data, evals and RL environments, built by experts on Scale’s platform and available to license off the shelf.
- 01
Built at the frontier
Created with the same quality standards Scale applies to its public benchmarks, and kept separate from benchmark test sets. Built by experts on Scale's platform.
- 02
Targeted at real gaps
Every dataset is designed around a specific, measurable model weakness — not generic volume. Tasks are curated to the cases where state-of-the-art models still fail.
- 03
Ready to license
Licensed off the shelf and delivered within weeks rather than quarters, Extend any dataset with custom tasks, extra domains or new locales.
Agentic Tool Use & RL Environments
Agents that look strong on short tasks often break on long ones. They lose track across dozens of tool calls, act on incomplete information or stall when systems disagree. These datasets and environments train and test agents on long, multi-tool work inside realistic enterprise systems.
Agentic tool use
Human-in-Loop (HiL)
Teaches agents to spot missing or conflicting information and ask the right clarifying question before acting. Focused on moments where a wrong assumption costs more than a short pause.
Agentic tool use
MCP Advanced
The hardest real-world tool-use tasks, built on real business cases. Each requires 50–100 tool calls and tests whether an agent can plan, recover from errors and finish the job.
Computer Use & GUI Trajectories
Computer-use agents have to read a screen, find the right control and recover when the interface shifts. These trajectories capture professionals doing real work in real desktop software, step by step, so agents learn every action a task takes.
CUA Professional Office
Computer-use trajectories across the office apps professionals rely on every day, captured end to end.
Safety, Red-Team & Frontier Risk
Safety teams need to see how models behave in adversarial and high-stakes settings. These datasets help labs measure risky behavior, track it across model versions and close gaps before release.
Cyber Pentest
Realistic penetration-testing tasks to measure and track offensive cyber capability before release. Gives safety teams a consistent yardstick across model versions. Available to qualified organizations.
Software Engineering & Coding Agents
Coding agents must set up their environment, run tests, read errors and try again. These datasets put agents in that loop, with automated verifiers that check whether the work actually runs.
Terminal Bench
Hard command-line tasks run in isolated containers. Automated verifiers check each one, so results are reproducible.
Coding
SWE Atlas
Real repo tasks across many programming languages. Each comes with environment setup and automated verifiers, ready for training and evaluation.
Coding
SWEAP Pro
Real engineering problems in real repos. Agents read the codebase, edit the code, run the tests and ship a working fix.
Coding
Tau 3
Next-gen tool-agent-user tasks with structured rubrics that grade both the final outcome and the steps the agent took to get there.
Multimodal: Vision, Document & Audio
These datasets pair images, charts, documents, video and live audio with tasks that need careful, multi-step reasoning. Domain experts write and check every task.
Humanity’s Sixth Sense
Tests intuitive visual reasoning across a wide range of image and video inputs: the snap judgments people make at a glance that models still get wrong.
Multimodal
Audio Full Duplex Tool Use
Real-time spoken conversations where the model calls tools mid-dialogue without breaking the flow. Built for voice agents that need to listen, act and respond at once.
Multimodal
Charts Reasoning
Scientific, financial and technical charts paired with expert-written answers. Questions require reading values precisely, comparing trends and drawing conclusions the data supports.
Expert Domain & Professional Judgment
Valuable professional work takes judgment, trade-offs and a finished product a client would accept. These datasets capture that work from practicing experts, so models learn what good looks like.
RLI
Training AI agents on real-world, expert-grade freelance work across 6 RLI domains and hillclimbing Scale's RLI. Every domain involves creative and professional judgment, spanning graphic design, audio and music, animation, video production, game development, and interactive web work (much of which is UI/UX).
Don’t see what you’re looking for?
This is a preview of the catalog. More datasets and environments are in production now across agents, coding, multimodal, safety, and more, and we can extend any existing dataset to fit your model’s gaps.
Tell us what you’re building and we’ll point you to the right fit.