Datasets:

---

nvidia

/

aisimulate-fpm-dataset

Like 9

Follow

NVIDIA 72.9k

Changes

• Use serving.layout: agg | disagg | unknown | not_applicable, independent of collection type. Self-benchmarks can use either deployment layout; missing evidence stays unknown. Remove the automatic self-benchmark-to-not-applicable mapping (15 runs corrected), and rename five confirmed disaggregated runs. Worker roles and topology remain independent.

• Migrate all 39 current and historical measurement snapshots to measurement manifest v5, configuration manifest v4, and catalog index v6.

• Record collection type, AgentX benchmark preset, replay mode, dataset, workload, serving layout, and immutable sources per collection run, bound to the actual truth files and logical record ranges. Keep unresolved metadata explicitly unknown.

• Move bulk iterations, reduced measurements, indexes, and request metrics to typed Parquet with Zstd compression. Preserve measurement values, rank grouping, exclusions, and evaluation order. Keep manifests/configuration in JSON and original compressed logs for traceability.

• Recover capture-matched request metrics and record unavailable evidence. Normalize canonical filenames while retaining original names in source references and raw archives.

• Update importers, schemas, catalogs, documentation, and the migration audit.

• Use descriptive asset names: fixed concurrency (iterations-c16-decode.parquet), min/max ranges (iterations-c16-128-unknown-decode.parquet, with unknown for unattributed records), and numbered fallback names when settings or units are unknown. Preserve provenance purpose and explicit settings (provenance/c64/agx_windows_decode.tsv). Same-name assets receive suffixes without overwriting. Stable IDs, bytes, ordering, and run bindings remain unchanged; references, catalogs, and audits are refreshed.

Validation

• Full dataset validation passed: 39 configuration snapshots and 291 measurement files, including checksums, catalog references, and collection bindings.

• Storage, request-recovery, grouped-measurement, self-benchmark, and DSV4.1 tests: 104 passed.

• AISim's migration comparison matched ordered measurements across 39 snapshots and 3,185,145 accepted observations. No predictor or evaluation-membership changes are intended.

• Pinned evaluation and dashboard preview use revision 562d9580, before the serving-layout metadata update. The new update preserves measurement files, IDs, order, and collection bindings; full dataset validation and serving-layout browser checks passed. The native evaluation has not been rerun for this metadata-only revision.

• Verified unchanged content hashes, file IDs, ordering, and run bindings across all 39 snapshots after renaming. A second refresh produces no byte changes.

Rollout

• Depends on AISimulate #417, including the Parquet logical-ID and optional-provenance fixes in 714e0188 and serving-layout display compatibility in 87ab89d3.
• Merge compatible AISim consumers first, then this dataset PR, then refresh evaluation and production Pages. Older pinned snapshots remain readable by the compatible consumers; historical evaluation artifacts remain unchanged.
• Source branch: simonec/fpm-manifest-parquet; reviewed dataset revision: e3a6c3fcbbe56102df960fb59ba5e97bb76c83f3.