Skip to content

Repository files navigation

vibe-ic benchmark data

Published measurement output from vibeic/vibe-ic, together with the exact design-input snapshots needed to interpret those results. The tooling, external benchmark corpus, and reusable IP live in their own repositories.

Layout

path what it is
evaluation/<benchmark>/input/ retained design inputs; not an evaluation result cell
evaluation/<benchmark>/v<version>_<llm>/ canonical open-benchmark result cells, keyed by benchmark, plugin version, and authoring model
ic/<design>/v<version>_<pdk>/ canonical IC campaign output, one directory per (design, plugin version, PDK)
protocol_parity/ deterministic Phase-1 protocol inputs and parity evidence; this is not an authoring-LLM benchmark result
PUBLISHING.md the publishing contract these directories must satisfy
BENCHMARK_IC_CAMPAIGN_STATUS.md campaign state at the time of the snapshot

What is here, and what is deliberately not

For IC physical-result cells, only cells that PASSED and conform to the layout are published. Three conditions, each measured rather than judged:

condition instrument
the directory is named v<version>_<PDK> PUBLISHING.md's layout rule
its RESULT.md says PASS the cell's own verdict
benchmark_evidence_structure_check.py exits 0 the repo's own validator

Four cells meet all three. Everything else was removed rather than shipped: a caravel_user_project cell whose own RESULT.md says FAIL and which fails the structure check; every clean_run_* directory; and every loose phase1/, phase2/, phase3/, reports/, sim/, steps/ tree that sat directly under an IC instead of inside a cell. PUBLISHING.md allows exactly input/ and v*/ at that level.

A published IC result that did not pass is not evidence of convergence, and mixing it with converged cells makes the passing ones harder to trust.

Open-benchmark results follow a separate retention rule described in evaluation/README.md: retain only the newest documented complete result for each (benchmark, Vibe-IC version, authoring LLM) identity. A result without an attributable plugin version or model is not placed in a named cell. The removed snapshot paths remain recoverable from the source snapshot commit recorded in SNAPSHOT.txt.

Where the inputs are

Design inputs are retained beside their result families, but outside version/model result cells:

  • IC inputs: ic/<design>/input/
  • Open-benchmark inputs retained from source runs: evaluation/<benchmark>/input/
  • Protocol inputs: protocol_parity/<protocol>/input/

External benchmark material lives in vibeic/benchmark-external; reusable IP and its pinned submodule commits live in vibeic/IP. Neither sibling repository is folded into this results repository.

What moved here on 2026-08-17

vibeic/vibe-ic now holds only the plugin. Three trees that were never part of it came here instead, so that their growth stops being the plugin's problem:

tree what it is
ic/, evaluation/ benchmark results, and the design inputs that produced them

Two sibling trees got their own repositories rather than being folded in here, because a results repository that also holds external material and IP pointers is three things wearing one name:

repository what it is
vibeic/benchmark-external external benchmark material
vibeic/IP the four IP submodule pointers and their pinned commits

These repositories are now the canonical homes, so future data growth is separate from the plugin. Historical copies and paths in the original vibe-ic history remain useful for provenance, but are not the publishing target.

The design inputs moved with their result families rather than being discarded: an input is the one part of a benchmark run that cannot be regenerated from the result artifact.

Provenance

Snapshot of vibeic/vibe-ic at the commit recorded in SNAPSHOT.txt. History was deliberately not carried over — this is a fresh tree, not a filtered rewrite, so commit SHAs here do not correspond to anything in vibe-ic. To read how a result came to be, use vibe-ic's own history for the same path.

Historical dangling-symlink inventory

DANGLING_SYMLINKS.txt records 31 dangling symlinks that existed in the source tree at snapshot time. Their loose steps/ trees are not part of the current canonical result layout, so the current tracked tree contains zero symlinks. The inventory remains as migration provenance; it is not a claim that those paths still exist here.

About

Vibe-IC benchmark results: open-benchmark evaluations (VerilogEval / RTLLM / CVDP) and canonical IC campaign outputs. Design inputs stay in vibeic/vibe-ic.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages