Datasets:

---

nvidia

/

aisimulate-fpm-dataset

Like 2

Follow

NVIDIA 69.3k

Follow-up to #2.

Summary

• Remove the eight redundant per-node FPM samples and two redundant truth samples.
• Retain both complete raw rank streams for provenance and validator cross-checking.
• Materialize listener windows with exact decimal timestamp and clock-offset arithmetic so the declared anchor and strict 8 ms half-open boundary match the Gym adapter.
• Regenerate the canonical GLM iteration streams, manifests, catalogs, audit reports, and documentation.

This does not change the latency policy: all rank measurements remain authoritative and max_rank_wall_time remains the cached default.

Validation

• uvx ruff format --check scripts/manage_dataset.py
• uvx ruff check scripts/manage_dataset.py
• uv run --with pyarrow python scripts/manage_dataset.py validate
• Repeated materialization produced identical decode and prefill SHA-256 hashes.
• Merged FPM Gym adapter loaded both complete files: decode 91,163 usable synchronized observations; prefill 167 usable observations, with incomplete and zero-latency groups classified as designed.