Datasets:
Commit History
-
fix(dataset): correct eight unsupported IFC-Bench reference answers
Qichen Liu 刘启辰Replace unsupported estimates with model-grounded values or explicit limitations across 4351, dental_clinic, hitos, schependomlaan, and sixty5.\n\nPreserve row identities and splits; only ground_truth fields change.d3bd2f6 -
fix(dataset): correct four IFC-Bench v2 reference definitions
Qichen Liu 刘启辰- mark UFA as undefined without a model-grounded usable-area basis - define the facade area scope and correct its component total - map Hitos ground level to 1.etg and enumerate its 57 spaces - replace overlapping Level 1 buckets with exact stored function categoriesee0b33f
-
fix(dataset): correct 33 Molio oracle answers
Qichen Liu 刘启辰- recompute all 58 Molio QA pairs against the restored IFC and update the 33 answers that differ\n- preserve question text, ids, categories, and split identity while refreshing the V2 checksum8f81413
-
fix(dataset): restore complete Molio IFC source
Qichen Liu 刘启辰f8024d1
-
fix questions 519 and 520 (same window max and min width) (#3)
- fix questions 519 and 520 (same window max and min width) (78314bd46b04cb1f96d928809c5688dae2803b42) Co-authored-by: Joost Rehder <powerkrieger@users.noreply.huggingface.co>b81e490
-
Deduplicate v2: remove train-side copy of a cross-split duplicate
sylvain hellinFormer rows id 21 (train) and id 22 (test) were byte-identical (a door-type fire-rating question, project 4351, category 2), a train/test leak. Removed the train copy (id 21) and kept the test copy (id 22), so the frozen 514-question evaluation split is unchanged. Ids are stable identifiers, so 21 is now an intentional gap (no renumbering). The exact pre-dedup snapshot is preserved at tag v2.0-frozen-split. - ifc-bench-v2.csv: 1026 rows (was 1027); category 2 count now 568 - eval-split-hellin2026.csv: 514 test / 512 train (test set unchanged) - README: deduplication note4b37e5d -
Add stable id column and frozen eval split (eval-split-hellin2026)
sylvain hellin- id: 1..1027 in existing row order, added byte-preserving (only the id prefix per record; original quoting untouched) - eval-split-hellin2026.csv: fixed 514-question held-out test split by id (seed-free), the split used in Hellin et al. 2026 - README: document the fixed evaluation splitaa5e6a1
-
configure HF dataset viewer to use QA CSVs
sylvain hellinWithout an explicit `configs` block in the README frontmatter, the Hugging Face viewer auto-detected the PDFs under projects/*/document/ and displayed them as the dataset, hiding the actual benchmark. Point the viewer at questions/ifc-bench-v2.csv (default) and expose v1 as a secondary config so the landing page shows the real QA rows and enables search.fa301cf -
added 7 BIM models (Rioverside project) + updated bibtex for V2
sylvain hellinebacaf2 -
corrected wrong answer
sylvain hellin94be85d
-
updated credits
sylvain hellin372af21
-
updated license to account for GPL3 models
sylvain hellin5b01aa2
-
IFC-Bench V2.0.0: 21 projects, 1027 QA pairs, 37 BIM models
sylvain hellin9e47a95