Erised Labs / Products

Data for models
that do the work.

Expert-curated tasks, explicit constraints, and results you can inspect. Built for teams turning model capability into reliable execution.

Explore the collection

The collection

01 / Dataset & evaluation tasks

GPU engineering · Auto Research

GPU Kernel
Optimization

Seven dependent pipelines from real engineering workloads. Evaluate how an agent handles quantization, attention, routing and gradients, then measure complete-pipeline speed against a reference.

Quantized linear
Sequence modeling
Sparse attention
Loss & gradients
Mixture of experts

Tasks: 7
Workloads: 112
Evaluated on: H200

Explore product Download datacard

Technical datacard · 29 October 2026

From specification to signal

Know what
you’re measuring.

A useful task makes its assumptions visible. Inspect the workload, understand the correctness check, and read performance in context.

01 / Inspect

Start with the task.

Read the interfaces, tensor shapes, precision boundaries and workload cases. See what an agent must preserve before it can optimize.

02 / Evaluate

Check the evidence.

Review correctness and performance together, including missing submissions, slower results and the limits of the timing protocol.

03 / Adapt

Bring your workload.

Discuss the hardware, operator domains and constraints that matter to your team, and scope a task set around your research goals.

A different capability in mind?

Talk with us about data, environments and evaluation for your next project.

Talk to the lab