There is a newer version of the record available.

Published October 19, 2026 | Version 0.1.0

ContamBench v0.1.0: Stressor Suites, Labeled Contaminated Workloads, and Reference Measurements for Testing LLM Serving

  • 1. Rochester Institute of Technology

Description

Stressor suites, labeled contaminated workloads, and per-request reference measurements for testing open-weight LLM serving stacks.

  • D1, stressor suites: five LoopLLM suites with 831 benign/stressor pairs, Stage-A (isolated) and Stage-B (under serving) adequacy, and three control-stressor sets.
  • D2, labeled workloads: replayable CSV workloads at contamination ratios 0 to 0.15 with manifests and configurations.
  • D3, reference execution: 24 runs on four capacity-calibrated single-GPU vLLM deployments, 796,500 request rows, SLO budgets, vLLM scheduler telemetry, and two GuideLLM runs.

The record also contains a source snapshot of the contambench Python package (Apache-2.0), which fetches this record, reproduces the published tables, and runs campaigns against an endpoint you operate. Start with README.md and DATASET_CARD.md.

The stressors raise serving cost by design. Run them only against systems you operate or are authorized to test.

Notes

Benign prompts derive from Stanford Alpaca (CC BY-NC 4.0). Model outputs remain subject to the license of the model that produced them.

Files

contambench-src-v0.1.1.zip

Files (627.7 MB)

Name Size
md5:b048bf64bdbc2b054845d0f676c73bc2
637 Bytes Download
md5:a881610b9b57b30c124965f335de1581
101.9 kB Preview Download
md5:669ca4cd14f52573dc6ee56cdb2ca699
422.1 kB Preview Download
md5:dcb37ed671c037ed727564e6a4f47a16
656.7 kB Preview Download
md5:355db663b27fd1d2ad082e6fe822fdc7
626.5 MB Preview Download
md5:e67a22394704e223cbc3f3eae41814aa
6.3 kB Preview Download
md5:058758a7f73a6d73561c22b8f6a6fcfc
1.7 kB Preview Download
md5:5d5ec8c66f725c57cf34351d050ae6a2
4.7 kB Preview Download
md5:6d154c036dd012c7b84c169f3db5a9ae
4.6 kB Preview Download
md5:ef216ad2c65da8dfac5573677fb68200
2.4 kB Preview Download
md5:97f1b936f901d1909db554fac804ccd8
821 Bytes Download