Publish expected_sets_sve.json: the definition set the SVE tables are scored against
Publishes the definition set the SVE result tables are scored against.
Why
Until now this set lived only in a local file on one machine. The dataset ships
every definition ever written, so anyone pulling it and scoring a model gets a
different denominator than the published tables use, with no error to warn them.
That is the whole problem this file fixes: pull the dataset, get the set.
What the numbers are
| dataset | n | how the set was arrived at |
|---|---|---|
| ncnn | 23 | 30 minus the 7 gemm definitions deleted in 650aa4b3 (they were gemv, not gemm) |
| llama.cpp | 26 | unchanged |
| kleidiai | 10 | full set -- see the caveat below |
| simd-loop | 26 | 47 -> 30 (those with a baseline-sve solution) -> 27 -> 26 |
simd-loop lost four definitions to scoring decisions, each because a baseline
was unrepresentative, not because the problem was bad:
loop_122-- reference isstd::sort(introsort); the hand-written SVE sorting
network is 964x slower, so the definition tests the wrong thing.loop_121-- baseline degrades to O(n^2) on duplicate keys.loop_032-- baseline omits a term; 10/10 workloads wrong.loop_010-- baseline is a correct, vectorized full scan that does not early-exit.
With the generated inputs the answer is settled by index 3, so the score is
proportional to N rather than to anything the model did. It scored 21.7x under the
old workloads, ~190x under the current ones, and 13,641x on sme2. Removing it changes
the ranking of the 2026-09-23 nanobot models, so it was not a rounding-level call.
Two things a reader should know
This file is SVE-scoped, and the name says so. sve2 and sme2 need their own sets:
the simd-loop baselines cover 30 / 36 / 24 definitions respectively, and the exclusions
above are properties of the SVE baselines, so they do not transfer unexamined.
The kleidiai 10 is the full set, not the SVE-runnable set. Only 3 of the 10 run on
Graviton3 -- gemm_fp32_n512_k512, gemm_qs4c32_n256_k512,conv2d_fp32_nhwc_kh3_kw3_s1_p1_c32_c64. The other 7 abort with SIGILL because their
KleidiAI micro-kernels use instructions the hardware lacks (verified by runningbaseline-kleidiai-arm on all of them: 3/3 pass, 3/3 sampled from the rest crash).
The 10 is kept here because that is what the published tables score against -- a
kleidiai cell shows 3/10 and is reported as incomplete rather than silently scored on 3.
Anyone reporting a kleidiai number on SVE should label it kleidiai-sve (3).
No definitions, solutions or runs are deleted by this PR; the excluded definitions stay
in the dataset because other ISAs still use them.
Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com