Datasets and environments that give frontier models domain expertise
Snorkel builds the human expert-authored datasets, evaluation environments, and benchmarks calibrated to push the limits of frontier model capability. Off-the-shelf or custom.
Where generic data runs out
Frontier model development stalls on data problems generic pipelines weren't built to solve, including distributional gaps in specialized domains, benchmark blind spots, and failure modes that only surface at scale. We build the data to solve them.
get started
Two ways to get the data you need
The data frontier models need most is rarely the data that already exists. Snorkel delivers it two ways: off-the-shelf for well-defined task areas, or custom-built for the gaps only you can see.
Curriculum-structured datasets for the task areas frontier models are pushing hardest, with rubrics, reviewer guidance, difficulty tiers, and eval slices built in.
Bespoke datasets, evaluation environments, and benchmark expansions to target the exact failure surface you're trying to close.
Datasets and environments
SWE-Bench Pro, Senior SWE-bench
Featured data series
Senior SWE-bench+
Real-world PR tasks at a senior engineer’s bar — inferring intent, runtime debugging, and shipping code that fits the codebase.
Enterprise agents
GDPval+
Economically-grounded professional tasks spanning all 20 O*NET sectors and 100+ occupations.
Frontier-Bench+
Multi-step terminal engineering tasks, calibrated to the Frontier-Bench standard.
Core benchmark contributor and reviewer
Closing gaps existing datasets can’t reach
Custom data development engagements start with the failure surface: what the model can't do, where it's brittle, and what the correct evaluation criteria are. From there, Snorkel builds the datasets, environments, and benchmark expansions needed to close it.
Research-validated methodology
