Datasets:

---

nvidia

/

aisimulate-fpm-dataset

Like 1

Follow

NVIDIA 69.1k

Moves the six published FPM tables from ai-dynamo/aisimulate#232 to the HF dataset, retaining the source-qualified runtime identities and calibration provenance.

• B300 TP2 and TP4, plus GB300 TP2; full and decoder_bounded remain distinct. Six tables, 145 rows each (870 total), with byte-identical Parquet payloads.
• These are calibration inputs, not independent truth. No private raw campaign archive is uploaded and no new accuracy or GPU collection result is claimed.
• GB300's historical worker shutdown failure and other qualification limits remain provenance; migration does not repair the collector or improve accuracy.
• Include explicit schema-7 ingestion and validation, system configurations, immutable source hashes and license notices. Do not borrow schema-6 execution identity.
• Source revision: 399f93cb875953719204c5ddf47804fea1fed835. AISimulate consumer will use the HF commit SHA and SHA-256 verification before the GitHub tables are removed.

The separate historical PR158 migration starts from the same dataset baseline. Refresh this contribution's catalogs after the first contribution merges; do not discard either contribution's entries when resolving shared catalog changes.

Request changes — reviewed commit 8c070411dfeba2b9211a448c67c8651c76431d27.

• [P2] Do not label the archive commit as the producer commit. The importer unconditionally sets aisim_commit=399f93cb875953719204c5ddf47804fea1fed835 and aisim_commit_status=known. This identifies the source archive; it does not establish the code that produced the measurements. Preserve recorded producer identities, use null with unknown/mixed status when no single producer commit is established, and retain the archive commit in source_revision. This shared importer issue also affects #11. Regenerate the affected manifests/catalogs and add a regression check.
• [P2] Retain collection provenance when removing evaluation metrics. The metadata projection drops the entire collection_provenance and private_publication_provenance sections. This removes full image digests, the recorded checkpoint revision fb2764a5cf321eaa5070ca8f9e892818f477c16d, source/runtime qualification, and the first-geometry-qualified-fixed-attempt selection policy. B300 also loses observed package versions and calibration attempt/qualification counts. The published manifest sets the known model revision to null. Project the collection/runtime fields into retained provenance while excluding prediction/error metrics, populate the known model revision, and validate their preservation. An original-sidecar hash identifies the source but does not carry these fields into the migrated dataset.

Validation passed: full dataset validation, 48 tests, byte-for-byte comparison of all six Parquet payloads (870 rows) against the pinned GitHub source, and all 30 file references in the AISimulate consumer manifest. GB300 worker-shutdown caveats remain intact. Native prediction and GPU accuracy were not rerun.

After fixes, merge #11 first and regenerate this PR's catalogs against merged main; the current heads conflict in five catalog files. Preserve both imports, rerun validation, and refresh affected consumer pins and hashes.

Addressed the requested changes in c968a615d3c00cbf71b41515012925118d577e0e.

• Producer identity is now derived from the recorded aic_revision, separately from the archive source_revision. Original Git and installed-package identities are retained; snapshots with multiple producers use null/mixed, and missing or non-Git producer evidence uses null/unknown. The importer no longer substitutes the archive commit.
• Collection-only provenance now retains the recorded image digests, model checkpoint fb2764a5cf321eaa5070ca8f9e892818f477c16d, runtime/source qualification, observed packages, fixed-attempt selection policy and calibration counts. The known model revision is populated. Prediction/error metrics remain excluded.
• Regenerated manifests and catalogs; added regression tests for producer/archive separation, collection provenance preservation and rejecting inconsistent identities. Full dataset validation and all 52 tests pass. All 6 calibration Parquet payloads remain byte-identical. Remote readback verified SHA-256 for all 43 contributed files at this revision.

AISimulate consumer pins and metadata hashes are being refreshed to this revision. These checks establish migration integrity, not new GPU collection or independent accuracy qualification. Merge #11 first; after that merge, I will refresh #12 against the actual merged main and validate the combined catalogs before requesting its merge.

Re-review passed. Both requested findings are resolved in c968a615d3c00cbf71b41515012925118d577e0e: producer identity is separate from archive identity, and collection/runtime provenance plus the known model checkpoint are retained. I merged main after #11 and regenerated the combined catalogs in 4f2cd66f517e55746e6d158a541187cf00060c91. Full combined dataset validation and all 52 tests pass: 53 FPM files overall, retaining all seven historical #158 tables and all six #232 tables. All 13 DeepSeek V4.1 payloads match their pinned GitHub originals byte-for-byte. Current AISimulate consumer manifests pin the corrected revisions and their file hashes match. Data and sidecar files were unchanged by catalog conflict resolution. No remaining merge blockers found; merging. These checks establish migration integrity, not new GPU accuracy qualification.