Datasets:

---

sylvainHellin

/

ifc-bench

Like 20

Corrects four IFC-Bench v2 reference answers without changing question text, stable ids, model/project/category assignments, or the frozen evaluation split.

Corrections

• Q128 — UFA: replaces the unsupported 2,981.73 sq m value with an explicit incomplete-information answer. The referenced structural IFC contains no IfcSpace entities and no stored UFA/usable-area basis.
• Q161 — facade area: defines the scope as explicit exterior vertical facade elements. The 100 IfcWall objects with Pset_WallCommon.IsExternal = true contribute 4,565.49305 sq m of Qto_WallBaseQuantities.GrossSideArea; the separate IfcCurtainWall contributes 18.0 sq m, for 4,583.49305 sq m total.
• Q560 — ground level: maps the Norwegian ground floor to 1.etg, not u.etg, and lists all 57 spaces with their rounded NetFootprintArea values. The unrounded total is 2,753.7635425 sq m.
• Q948 — Level 1 breakdown: replaces overlapping synthetic buckets whose displayed percentages totaled 127% with the 17 mutually exclusive stored PSet_Revit_Other.Category Description values. The 60 spaces total 1,800.849847 sq m; raw percentages total 100% and independently rounded values total 99.97%.

Validation

• parsed all 1,026 v2 rows and confirmed only the four ground_truth cells changed
• preserved the frozen split exactly: 514 test / 512 train
• recomputed each quantity, room list, group count, area, and percentage from the pinned IFC files with IfcOpenShell 0.8.3.post2

Thanks, Q128, Q560, and Q948 reproduce exactly. Q161 needs one clarification.

The 100 external IfcWall objects sum to 4,565.49305 m² of GrossSideArea, which checks out. The additional 18.0 m² for the IfcCurtainWall has no stored quantity, and I traced where it comes from: the curtain wall element has a NULL representation, and 18.0 is exactly the 6.0 m x 3.0 m elevation bounding box of its decomposed members. The actual meshed glazed-panel area is 13.4 m² per side, not 18.0.

Reproducible check:

import ifcopenshell, ifcopenshell.geom as geom, numpy as np
from ifcopenshell.util.element import get_psets
f = ifcopenshell.open("projects/digital_hub/arc.ifc")
ext = [w for w in f.by_type("IfcWall")
       if get_psets(w).get("Pset_WallCommon", {}).get("IsExternal") is True]
tot = sum(p["GrossSideArea"] for w in ext
          for p in get_psets(w, qtos_only=True).values() if "GrossSideArea" in p)
print(len(ext), round(tot, 5))                     # 100  4565.49305
cw = f.by_type("IfcCurtainWall")[0]                 # NULL representation on the element itself
parts = [o for r in cw.IsDecomposedBy for o in r.RelatedObjects]
s = geom.settings(); s.set("use-world-coords", True)
V = np.vstack([np.array(geom.create_shape(s, p).geometry.verts).reshape(-1, 3) for p in parts])
bb = V.max(0) - V.min(0); d = sorted(bb)
print(round(d[-1] * d[-2], 3))                      # 18.0  (6.0 x 3.0 elevation bbox)

So 18.0 is an elevation-bbox estimate, not a stored or glazed-panel quantity. Two clean options: keep it but state in the answer that the curtain-wall term is a 6x3 elevation bounding box (giving 4,583.49), or drop it and report only the stored-quantity total of 4,565.49. I would prefer to keep it with the calculation explanation, but both are fine, just pick one and note the basis.

Thanks for tracing that. I kept the 4,583.49 m² total, but rewrote Q161 so the basis is explicit: 4,565.49 m² comes from the stored GrossSideArea values, while the curtain-wall contribution is a 6.0 m × 3.0 m elevation bounding-box estimate because that element has no stored area quantity. The answer now also says that the 18.0 m² term is not a meshed surface-area measurement.

I pushed the update in 311b076 and refreshed the CSV checksum. Thanks for the careful review.

I’ve folded these four corrections into #8 alongside the Molio repair from #4. The replacement is based on the current main, keeps the clarified Q161 calculation basis, and has no merge conflicts. Closing this one so you only have a single clean PR to merge. Thanks for tracing the curtain-wall contribution and helping make the reference answer precise.