Benchmarks
Measure performance on real business tasks with defined inputs, expected outcomes, and scoring criteria.
FOR FRONTIER LABS
Train and evaluate
on real work.
Build model capabilities with curated datasets and authorized enterprise workflows. We shape useful tasks, trajectories, and outcomes into data for training and evaluation.
01 / DATA PRODUCTS
From a capability goal
to a defined deliverable.
Specify the domain, task format, difficulty, and scale.
We align the data and review criteria to your brief.
EVALUATE
SFT datasets
Learn from selected and reviewed examples, including task execution, tool use, corrections, and successful outcomes.
RL environments
Practice in reproducible task environments with tests, graders, or other verifiers that measure the result.
02 / WORKFLOW DATA
Capture the task.
Preserve the learning signal.
An execution log becomes useful data through selection, transformation, and verification.
- 01
Agree on scope and permitted uses.
Define eligible workflows, collected fields, processing boundaries, and the intended training or evaluation use with the data provider.
- 02
Select the useful parts of the work.
Identify task context, tool actions, corrections, and outcomes in approved workflows, such as coding-agent session trajectories.
- 03
Transform sensitive content and verify quality.
Remove or transform sensitive details. Where needed, reconstruct tasks and environments, then review examples and checks against the agreed specification.
- 04
Deliver for the authorized purpose.
Package approved benchmarks, SFT datasets, or executable environments with their task specifications, evaluation criteria, and agreed usage scope.
03 / ILLUSTRATIVE WORKFLOW
One debugging task.
Three possible products.
A coding session may contain a task, tool calls, failed attempts, a correction, and a passing test. The suitable product depends on data quality and the approved use.
APPROVED CODING SESSION
Task → Investigation → Patch → Test result
- Benchmark
- A held-out debugging task with tests and scoring criteria.
- SFT dataset
- A reviewed solution trajectory for learning the task.
- RL environment
- A reproducible repository state, task, and executable verifier.
Illustrative example. Evaluation and training uses are defined separately for each project.