alei97/VLMDRC-ASAP7-eval

VLMDRC — ASAP7 DRC-Understanding Evaluation Dataset

Evaluation crops for Vision-Language Model design-rule-checking (DRC) at an advanced (7nm, ASAP7) node. Each sample is a rendered layout crop labeled with the Calibre-signoff DRC violations it contains (rule name + exact geometry), for benchmarking whether a VLM can identify which design rule is violated, where, and why.

Companion code (rendering, eval, metrics, reasoning-trace collection): https://github.com/glzhou97/VLM_DRC_ASAP7

Source & provenance

• Layouts: OpenROAD-flow-scripts GCD benchmark routed on the open ASAP7 PDK (asap7sc7p5t), plus model-augmented variants.
• Labels: Mentor Calibre DRC (ASAP7 rule deck) — exact violation geometry (edges for spacing, polygons for enclosure/width).
• Scope: local interconnect — M1, M2, V1 (paper representation).
• License: ASAP7 is BSD-3-Clause; all layouts are open benchmarks (no proprietary foundry data).

Splits

Folder	N	Description
table2_set/	100	Benchmark: single-violation crops, 10 rules, + via_count, via_centers_um, measured geometry_text
table2_set_dim/	100	Same crops with measurements drawn on the image (via sizes, spacing dimension lines)
eval_large/	250	Larger single-violation set (verified Calibre geometry: edges_um, polygon_um, region_um)
eval_large_loc/	250	Same, with the exact Calibre violation region marked (localization oracle; no rule name)
real_annotated/ synthetic_annotated/	100 each	Violation-region overlay + layer legend + rule list (human-inspectable)
real/ synthetic/	100 each	Clean crops (real OpenROAD DRVs / model-augmented DRVs)

Format

Each split has images/*.png + eval_manifest.json (list of samples):

{
  "sample_id": "t2_000", "image": "images/t2_000.png",
  "source_gds": "6_final.gds", "crop_origin_um": [x, y], "crop_size_um": 0.256,
  "via_count": 1, "via_centers_um": [[x, y]],
  "violated_rules": ["M1.S.2"],
  "violations": [{"rule": "M1.S.2", "type": "edge",
                  "edges_um": [[x1,y1,x2,y2], ...], "region_um": [x0,y0,x1,y1]}]
}

Rules covered (M1/M2/V1): M1.S.2/4/5/6 (spacing), V1.W.1 (width), V1.S.1 (via spacing), V1.M1.EN.1 / V1.M2.EN.2 (enclosure), V1.AUX.1 / V1.M2.AUX.2 (auxiliary enclosure).

Coordinates are micrometers; images have µm axes with a fine grid for reading geometry. group / group-conditioned-spacing rule categories from the source paper are N/A for ASAP7 (no group-via rules).

Baseline results (100-sample benchmark, k=1, rule-ID F1)

Model	F1	VIA Loc	Spacing F1
Qwen3-VL-8B (base)	16.3	57.5	0%
GPT-5.2	11.7	88.9	0%
Gemini-3.1-Pro	13.5	66.7	0%

No baseline detects spacing violations from the raw crop — the gap domain adaptation is meant to close.