Datasets:

---

BIM-Edit

/

BIM-Edit-Tasks

like 0

Thank you for publishing the task prompts so quickly — that was exactly what we needed.

While validating tasks.jsonl against the IFC models in BIM-Edit/BIM-Edit, we found that 7 of the 216 target.guids do not exist in either the input or the ground-truth model (209/216 = 96.8% resolve). Since the README states that guids is used to filter candidate matches during scoring, these tasks may not score correctly regardless of what the agent produces.

This PR fixes those 7 entries plus 4 prompt-level issues. After the change, 216/216 target.guids resolve (100%) and no GUID violates the 22-character IfcGloballyUniqueId length. Task order, task IDs and the key structure of every record are unchanged; 10 records differ.

1. target.guids that do not exist (7 records)

1a. Extra 5-character prefix (3 records)

1pUd$18rZd98AjEd9pV6_jMu$XT is 27 characters. The trailing 22 characters, 18rZd98AjEd9pV6_jMu$XT, do exist in both 9d81a20e025d.ifc and 9d81a20e025d-update-ground-truth.ifc; the leading-22 variant does not. So the value appears to have picked up a spurious 1pUd$ prefix.

The prompt of WAL-UPD-DIR-R-002 also contains the 27-character ID, so that task currently asks the agent to rotate a wall by an ID that is not in the model. The prompt is corrected accordingly. -SPA- and -TOP- reference the target indirectly with valid GUIDs, so only their metadata changes.

1b. Counter overrun (4 records)

These GUIDs look like the result of incrementing a string past the last value that actually exists. Note that z9 → z10 also grows to 23 characters, breaking the 22-character IfcGloballyUniqueId length.

The replacement was recovered from the input↔ground-truth delta — the same fallback the README already prescribes for create tasks. In each case the delta contains exactly one removed element of the declared entity_type:

For WIN-DEL-*-A-002 the delta removes 14 entities, exactly one of which is an IFCWINDOW. For DOR-DEL-*-A-002 it removes 12, exactly one of which is an IFCDOOR. The -SPA- and -TOP- variants of each pair share the same input/ground-truth models, so they legitimately share a target.

The prompts of all four describe the target indirectly and use valid reference GUIDs, so the prompts are left untouched.

2. Prompt-level issues (4 records)

These change benchmark content rather than metadata, so please review them separately — we are happy to drop this section if you would rather make the edits yourself.

How this was checked

All 324 tasks were resolved against the models in BIM-Edit/BIM-Edit, mapping data/artificial/* → simple/* and data/realistic/* → complex/*. GUIDs were collected from the first argument of every IFC entity instance. Before: 209/216 resolve, 5 GUIDs are not 22 characters. After: 216/216 resolve, 0 length violations.

Two unrelated notes

• complex/1c98aa6c7ea5-create-input.ifc.ifc has a doubled extension; tasks.jsonl refers to it as 1c98aa6c7ea5-create-input.ifc. That is the only file whose name does not line up.
• Appendix F.5 says every reported score can be verified from cached per-task evaluation reports in runs/, but BIM-Edit-runs currently contains only the edited IFC outputs for two of the seven models, with no evaluation reports. Are the evaluator and the reports something you plan to publish? Being able to reproduce the reported numbers from the released outputs would help a lot.