jasminexli/verifier-challenge-traces

Verifier Challenge Traces v0.1 是一个包含真实2×H100工作负载的GPU遥测数据集，带有诚实和对抗性阶段。该数据集是Inspector Agents模式层原型的配套数据集，旨在解决翻译层差距问题——即使证明者的标签存在欺骗，验证者能否从原始遥测数据中推断出实际运行的内容。v0.1版本是一个标签翻转基准，对抗性阶段扰动声称的标签但不改变原始遥测数据。数据集包括多个阶段的GPU功率、利用率、内存、时钟和温度数据，以及NCCL日志和网络设备日志。数据集布局包括每个阶段的CSV文件和日志文件，以及一个包含主机/GPU/python/git/时钟信息的provenance.json文件。数据集还提供了阶段时间表、评估协议（留一主机出法）和注意事项，如主机数量有限、单节点NVLink等。

Verifier Challenge Traces v0.1 is a dataset containing ground-truth-labeled GPU telemetry from real 2×H100 workloads, with both honest and adversarial phases. It serves as a companion dataset to the Inspector Agents schema-layer prototype, addressing the translator-layer gap—enabling verifiers to infer actual operations from raw telemetry even when the provers labels are deceptive. The v0.1 version is a label-flip benchmark where adversarial phases perturb the claimed labels without altering the raw telemetry. The dataset includes GPU power, utilization, memory, clocks, and temperature data across various phases, along with NCCL logs and network device logs. The dataset layout features CSV and log files per phase, plus a provenance.json file with host/GPU/python/git/clock details. It also provides a phase schedule, evaluation protocol (leave-one-host-out), and caveats such as limited host count and single-node NVLink constraints.