Datasets:

---

nvidia

/

aisimulate-fpm-dataset

Like 2

Follow

NVIDIA 69.3k

Summary

• Move DSV4-Pro benchmark-lane configs from measurements/lanes/ to fpm/provenance/ and link them from matching FPM metadata.
• Flatten DSV4-Pro derived-truth CSVs under measurements/ and move serving-run helpers to measurements/provenance/.
• Apply the layout consistently to current and historical snapshots while preserving measurement and snapshot IDs.
• Refresh catalogs, manifests, hashes, audits, documentation, and validation.

Validation

• Dataset validation: 17 current configurations, 7 history snapshots, 37 FPM files, 175 measurement artifacts, 18 derived-truth files, and 13 FPM provenance files.
• All 24 current and historical measurement cases load through FPM Gym.
• FPM Gym dataset tests: 26 passed.
• JSON Schema validation passed for all 24 configuration and 24 measurement manifests.
• Ruff format/check and git diff --check passed.

Extended this PR to normalize every model and historical snapshot:

• truth and derived_truth files are direct children of measurements/.
• Configuration, window, and helper-manifest evidence is under measurements/provenance/.
• Only explicitly linked FPM producer configs remain under fpm/provenance/.
• MiniMax B200 duplicate truth basenames use concurrency-qualified names such as c16-fpm_stream_decode.jsonl.gz.

All 175 measurement IDs and payload hashes are preserved. Validation covered all 24 manifests, all 24 FPM Gym cases, and the 26 offline HF dataset tests.