Datasets:

---

sylvainHellin

/

ifc-bench

Like 22

Hi, I think just changing the file is not sufficient to fully address the issue with Molio.
For example, on question

608,"What are the key building metrics including total elements, number of spaces, and building levels with their elevations?"

it would appear to me that now the valid answer should not contain

Total Spaces: 108 defined spaces

as from a quick search in the file it contains 117 spaces instead, when using the original / uncorrupted file. I did not verify the other questions related to this file, so there might be more issues created by the NUL bytes.

Thanks for catching this. I audited all 58 Molio questions (Q607–Q664) against the restored IFC rather than only replacing the source file.

The follow-up commit 5ea2bca corrects 33 oracle answers and leaves 25 unchanged. In particular, Q608 now reports 117 spaces. The audit also corrected the derived area ratio, MEP component count, window and railing facts, and several ceiling, furniture, metadata, type, and material answers.

I verified that question text, IDs, categories, and train/test splits are unchanged, and that the dataset still loads as 1,026 rows with exactly the expected 33 ground_truth changes.

Thanks for reporting this issue. The restored arc.ifc matches the authoritative buildingSMART LFS object (sha256 626654e3…), and it seems that 32 of the 33 corrected answers reproduce exactly against it, but one response (to Q608) needs a fix before merge.

Q608 reports:

3,341 physical building elements when IfcOpeningElement (245) and IfcBuildingElementProxy (2) are excluded from IfcBuildingElement.

IfcOpeningElement is not a subtype of IfcBuildingElement, so it was never in that count and must not be subtracted. The correct figure is 3,588 − 2 = 3,586 (or 3,588 if proxies are kept).

Reproducible check:

import ifcopenshell
f = ifcopenshell.open("arc.ifc")  # the clean Molio model
print(len(f.by_type("IfcBuildingElement")))        # 3588
print(len(f.by_type("IfcOpeningElement")))         # 245
print(len(f.by_type("IfcBuildingElementProxy")))   # 2
print(f.by_type("IfcOpeningElement")[0].is_a("IfcBuildingElement"))  # False

Could you update Q608 to 3,586 and adjust the "245 openings excluded" wording? Everything else looks good to merge.

You're right. I've corrected Q608 to 3,586 non-proxy IfcBuildingElement entities: 3,588 total, less the 2 IfcBuildingElementProxy entities. The answer now lists the 245 IfcOpeningElement entities separately and makes the class relationship explicit.

The fix is in 231ff8b, with the CSV checksum refreshed. Thanks for checking the IFC hierarchy so carefully. I’ll resolve the existing README/CSV conflicts against the latest main after the smaller reference-definition PR lands, so the final checksum only needs to be regenerated once.

I’ve folded this into #8 together with the reviewed reference-definition fixes from #6. The replacement is based on the current main, includes the Q608 correction from your review, and has no merge conflicts. Closing this one so there’s only a single clean merge path. Thanks again for checking the restored model and the IFC class hierarchy so carefully.