jasminexli/verifier-challenge-traces

Verifier Challenge Traces — v0.1

Ground-truth-labeled GPU telemetry from real 2×H100 workloads, with honest and adversarial phases. Companion dataset to the Inspector Agents schema-layer prototype: this dataset addresses the translator-layer gap — given raw telemetry from a prover, can a verifier infer what's actually running, even when the prover's labels lie?

v0.1 is a label-flip benchmark. Adversarial phases here perturb the claimed labels but not the raw telemetry. v0.2 will add translator-layer tampering (subsetted NCCL logs, fake-NCCL-during-inference, sub-second interleaved train/infer) — sketches in the source repo's followup-adv-tests.md.

Quickstart

The dataset is an artifact collection, not a single tabular table. The two HF-viewable configs are convenience accessors; most users will iterate over the per-phase directories directly.

from huggingface_hub import snapshot_download
local = snapshot_download("jasminexli/verifier-challenge-traces", repo_type="dataset")

# Or load just the per-second nvidia-smi power/util via the viewer config:
from datasets import load_dataset
nvsmi = load_dataset("jasminexli/verifier-challenge-traces", "nvsmi", split="train")
labels = load_dataset("jasminexli/verifier-challenge-traces", "labels", split="train")

Layout

Two hosts, each running the full 8-phase schedule independently and concurrently on RunPod 2×H100 SXM:

runs/<UTC_timestamp>-<runpod_pod_id>/
├── provenance.json           # host/GPU/python/git/clock at start AND end
├── workload_labels.jsonl     # one start/end record per phase (claimed vs truth)
├── workload.log              # orchestrator stdout
├── checkpoints/              # gpt_tiny.pt, gpt_small.pt
└── phases/<phase_id>/
    ├── nvsmi.csv             # 1 Hz GPU power/util/mem/clocks/temp (proper CSV)
    ├── dcgm.csv              # 10 Hz dcgmi dmon — see note below
    ├── nccl_<host>_<pid>.log.gz   # one per rank, NCCL_DEBUG=INFO COLL+INIT, gzipped
    ├── netdev.log.gz         # /proc/net/dev @ 1 Hz, gzipped
    ├── stdout.log
    └── stderr.log

Note on dcgm.csv: v0.1 wrote DCGM telemetry from dcgmi dmon to a .csv extension, but the format is actually space-separated text with repeating #Entity header rows (closer to top output). It is not standard CSV; HF's auto-viewer cannot parse it. Parse with whitespace splitting and skip lines starting with #Entity or ID. v0.2 will rename to dcgm.dmon.

Phase schedule

#	Phase ID	Duration	Truth op	Claimed op	Notes
1	honest_pretrain_tiny	30 min	training	training	gpt-tiny pretrain on TinyStories
2	honest_pretrain_small	30 min	training	training	gpt-small (GPT-2 Small) pretrain
3	honest_inference_small	15 min	inference	inference	loads phase-2 ckpt, no_grad gen
4	honest_finetune_tiny	20 min	training	training	gpt-tiny finetune on WikiText, frozen embed
5	idle	5 min	idle	idle	confound check
6	adv_train_as_infer	15 min	training	inference	P1 attack — workload identical to phase 2, label lies
7	adv_big_as_small	15 min	training	training	P2 attack — runs gpt-small, claims gpt-tiny
8	adv_finetune_as_pretrain	15 min	training	training	P3 attack — workload identical to phase 4, claims pretrain

Phases 6, 7, 8 are the lies. The lie lives in claimed.* of the corresponding workload_labels.jsonl start record; truth.* is what actually ran.

Evaluation protocol — leave-one-host-out

The whole point of running multiple hosts in parallel is to make cross-host generalization measurable. The recommended primary metric is leave-one-host-out (LOHO) AUROC, not pooled.

For each property:

1. Train on data from N-1 hosts (all phases included).
2. Test on the held-out host.
3. Repeat for every host; report mean AUROC and per-split spread.

Per-host evaluation (train and test on the same host) is trivial — the verifier wins by learning host fingerprint. The cross-host floor is the actual scientific bar. If a verifier's pooled AUROC is much higher than its LOHO AUROC, it's overfit to host fingerprint, and the gap is itself a useful number to publish.

Baseline (mean nvsmi power per phase + 4 other simple features → logistic regression, P1) achieves mean LOHO AUROC = 1.000 on this dataset — i.e. the simplest possible features trivially separate training from inference. This is the v0.1 floor; serious verifiers should target much harder evaluations and add adversarial perturbations on top.

Caveats

• n=2 hosts. Cross-host axis exists but variance estimates are weak. v0.2 will scale to ≥3 hosts in distinct RunPod regions for stronger generalization signal.
• Single-node, NVLink only. No multi-node fabric, no InfiniBand. NCCL traffic is NVLink P2P, not the topology a frontier verifier will encounter.
• Adversarial subset is ~25% of run time. Benchmark numbers should be class-balanced or report the imbalance.
• v0.1 adversarial phases mostly perturb workload/labels. Deeper translator-layer tampering (subsetted NCCL, fake-NCCL-during-inference, sub-second interleaved train/infer) is in the source repo's followup-adv-tests.md and will land in v0.2.
• DCGM availability is host-dependent. When the container is unprivileged, a dcgm.skipped marker file is written instead of dcgm.csv. v0.1 captured DCGM on both hosts, but downstream code should not assume it.
• Clock skew between hosts is recorded in provenance.json (clock_offset_start_us, clock_offset_end_us). Hosts whose offset shifted >1s during the run should be flagged unreliable for cross-host temporal alignment; v0.1 hosts were stable.

Source data not redistributed

The training corpora — TinyStories (CDLA-Sharing 1.0) and WikiText-2 (CC-BY-SA 4.0) — are not shipped in this dataset. Only the telemetry and labels generated from them are. Users who want to reproduce the exact inputs should re-download from HuggingFace (roneneldan/TinyStories, wikitext/wikitext-2-raw-v1).

Source repository

Code that produced this dataset (orchestrator, capture wrapper, configs, smoke test): https://github.com/jasonhausenloy/inspector-agents/tree/main/gpu-runs

Companion paper / live demo: https://jason.ml/inspector

License

Telemetry and labels: CC-BY-4.0. Users assume the licensing of any source corpora they re-download.