SoroushVahidi/llm-serving-common-contract-m3

Common-contract LLM-serving scheduler results — M3

Release m3-v1.0.0. Author: Soroush Vahidi. Promoted from independently verified rc3 with unchanged result-table bytes. This is the authoritative M3 table distribution.

This dataset provides saved result tables supporting Scheduler Equivalence and Performance Trade-offs under a Common LLM-Serving Contract, manuscript commit 42170fdcdc7c4b68ae2639b319a4892d097532a0. It is a table view of the complete reproducibility artifact, not a separate experimental campaign. The source repository is https://github.com/SoroushVahidi/llm-serving-scheduler-robustness-benchmark. Source release: https://github.com/SoroushVahidi/llm-serving-scheduler-robustness-benchmark/releases/tag/m3-v1.0.0. Hugging Face table distribution: https://huggingface.co/datasets/SoroushVahidi/llm-serving-common-contract-m3. Kaggle mirror: https://www.kaggle.com/datasets/soroushvahidi/llm-serving-common-contract-m3. The complete saved-evidence archive is published at https://zenodo.org/records/23608723 (DOI: https://doi.org/10.5281/zenodo.23608723). All four destinations retain their distinct roles under m3-v1.0.0.

The four-policy CM-0.1 CPU simulation contains 5,760 evaluation cells (120 windows × six settings × two capacities × four policies), including 144 reused pilot cells and 5,616 additional executions. Ranking and pairwise tables contain 15,840 and 95,040 rows. The 7,200 STAT01 endpoint summaries describe 134,000 saved synthetic studies, including reused records; they are not 7,200 new experiments. Budget sensitivity contains 72 cells and urgency checks contain 18 certificate rows. All 33 table counts and byte identities are in TABLE_SCHEMA.json.

Windows derive from Azure, Bailian/Qwen and BurstGPT. This table-only distribution excludes request-level traces, anonymous session/cache identifiers, bulk event journals, latent statistical arrays and the older LSSP matrix. Source acquisition and exact third-party terms accompany the complete canonical archive. Microsoft Azure, Alibaba Cloud and the BurstGPT authors retain their upstream rights. Eligible original result tables and this documentation are CC BY 4.0; the included verifier retains MIT. No raw trace rights are granted.

The dataset is separately versioned from SoroushVahidi/llm-serving-scheduler-portability, whose Phase-12 campaign has 18,720 executions/9,360 configurations and a different contract/policy panel. All 720 M3 source-window-setting labels share prior input lineage; do not treat shared windows or mirrored tables as independent observations. Older DOI 10.5281/zenodo.22306798 identifies LSSP software, not this dataset. See DUPLICATION_ASSESSMENT.json.

Use DATA_DICTIONARY.md and TABLE_SCHEMA.json before analysis. Missing values remain missing. Interpret conditional latency with completion counts, distinguish means of window p95 from pooled p95, and retain descriptive reversal denominators. C01 comparisons are finite-corpus descriptions, not population inference. Statistical calibration families distinguish prespecified from postdiagnosis procedures. SLO-gated improvements are not evidence of urgency activation. No native serving-system or GPU-performance validation is claimed. TIME01 and ABL01 are included after independent scientific review and author-authorized integration. TIME01 contains 384 new variable-duration executions and 384 reused fixed baselines across 24 purposively selected windows, two settings, two capacities and four adaptations. ABL01 has 1,440 paired original/derived controls and zero new simulations; its frozen simulated:false identities are preserved. These dependent views are not additional independent experiments. TIME01 is synthetic, not hardware-calibrated; selected BurstGPT mean latency is worse under both clocks. ABL01 isolates the prompt-ordering trade-off within the original nonbinding contract; it does not establish universal optimality. Historical frozen scientific_review_required:true flags remain unchanged and are superseded by the dated hash-linked independent review verdict, not silently edited.

HF and Kaggle use identical CSV bytes under m3-v1.0.0; native platform revisions do not define new scientific data. Verify with sha256sum -c SHA256SUMS and python verify_tables.py .. For saved-record reproduction, obtain the complete canonical artifact from Zenodo via https://doi.org/10.5281/zenodo.23608723. The immutable archive preserves rc3 preparation metadata as provenance; current publication links and status are supplied by the public record. The BOUND01 proof belongs to the complete archive, not these numeric tables.

Example offline loading:

import csv
with open("c01-all-cells.csv", newline="") as f:
    rows = list(csv.DictReader(f))
assert len(rows) == 5760
assert len({tuple(r[k] for k in ("source", "window", "region", "profile", "policy")) for r in rows}) == 5760

Upstream citations: Stojkovic et al., DynamoLLM, https://doi.org/10.1109/HPCA61900.2025.00102; Wang et al., KVCache Cache in the Wild, https://www.usenix.org/conference/atc25/presentation/wang-jiahao; BurstGPT, https://doi.org/10.1145/3711896.3737413. No provider endorsement is implied.

The Hugging Face split name train is a loader convention, not a training/test partition. Use the immutable m3-v1.0.0 revision after publication.