README.md

# BIM-Edit Tasks

324 natural-language building model editing tasks over IFC (Industry Foundation Classes) models.
Each task pairs an input IFC file with a ground-truth IFC file and a natural-language instruction
describing the edit that turns one into the other. The benchmark measures whether an agent can
translate a spoken-language design instruction into a correct geometric and semantic change to a
building model.

## Composition

The 324 tasks are fully balanced across three axes.

| Axis | Values | Count each |
|---|---|---|
| Operation | `create`, `update`, `delete` | 108 |
| Category | `direct`, `spatial`, `topological` | 108 |
| Scene | realistic (`scene:R`), artificial (`scene:A`) | 162 |

Element types covered:

| IFC class | Tasks |
|---|---|
| `IfcWall` | 72 |
| `IfcSlab` | 72 |
| `IfcSpace` | 72 |
| `IfcDoor` | 36 |
| `IfcWindow` | 36 |
| `IfcColumn` | 36 |

The tasks draw on 88 distinct input models and 98 distinct ground-truth models.

### Categories

- **direct**: the instruction states the edit explicitly, including coordinates, dimensions and the
  identifiers of the elements to connect to. No inference about the surrounding model is needed
  beyond locating the named entities.
- **spatial**: the target is described relative to other geometry ("0.9m away in -y direction from
  the stairway"). The agent must first locate the reference element and then derive the placement.
- **topological**: the instruction is phrased in terms of relationships (containment, bounding,
  connectivity, openings) rather than coordinates.

### Scenes

- **realistic** (`data/realistic`): models derived from real building projects, with the geometric
  irregularity, redundant entities and inconsistent property sets that come with them.
- **artificial** (`data/artificial`): synthetic models built to isolate a single edit, with clean
  and predictable structure.

## Fields

Each line of `tasks.jsonl` is one task object.

| Field | Type | Description |
|---|---|---|
| `task_id` | string | Unique identifier, see the format below. |
| `operation` | string | One of `create`, `update`, `delete`. |
| `category` | string | One of `direct`, `spatial`, `topological`. |
| `input_ifc` | string | Relative path to the IFC model handed to the agent. |
| `ground_truth_ifc` | string | Relative path to the reference IFC model after the edit. |
| `prompt` | string | The natural-language instruction. |
| `target` | object | `{"entity_type": <IFC class>, "guids": [<GlobalId>, ...]}`. |
| `evaluation` | object | Per-task scoring overrides. Empty for every task in this release, meaning all tasks use the default evaluator settings. |
| `tags` | list[string] | Operation, IFC class, `scene:R` or `scene:A`, category, and `anchored` or `no_anchor`. |

### `target`

`entity_type` is the IFC class of the element being edited and is used to filter candidate matches
during scoring. `guids` lists the GlobalIds of the affected elements in the ground-truth model. It
is populated for all 108 `update` and all 108 `delete` tasks and is empty for the 108 `create`
tasks, where the created element does not exist in the input model and is instead recovered from
the delta between `input_ifc` and `ground_truth_ifc`.

### `anchored` vs `no_anchor`

216 tasks are `anchored`: the prompt refers to at least one existing element that fixes the edit in
place. 108 are `no_anchor`, giving absolute coordinates instead. This tag separates tasks that test
model comprehension from those that test instruction following alone.

### `task_id` format

`<ELEMENT>-<OPERATION>-<CATEGORY>-<SCENE>-<INDEX>`, for example `WAL-CRE-DIR-R-001`.

| Segment | Values |
|---|---|
| Element | `WAL` wall, `DOR` door, `WIN` window, `SLB` slab, `ROM` space, `COL` column |
| Operation | `CRE` create, `UPD` update, `DEL` delete |
| Category | `DIR` direct, `SPA` spatial, `TOP` topological |
| Scene | `R` realistic, `A` artificial |
| Index | Zero-padded counter within the group |

## Example

```json
{
  "task_id": "WAL-CRE-DIR-R-001",
  "operation": "create",
  "category": "direct",
  "input_ifc": "data/realistic/0e29802b440f-input-create_wall.ifc",
  "ground_truth_ifc": "data/realistic/0e29802b440f.ifc",
  "prompt": "Create a wall with a length of 2.62m from (x1,y1,z1) = (7.180, 14.140, 3.000) to (x2,y2,z2) = (9.800, 14.140, 3.000) with a thickness of 0.240m in -y direction and a height of 3m. Make sure that the wall is added to the building storey with id 1jPCssxb5C1RSM_YHIx$zl, connected to walls 30a7nM35T5pgZzaItbPb1u and 153QDldl9AdhGy6O1dePed, and bounds the spaces 19QUlaWcT26g1KZHffEeW9 and 19QUlaWcT26g1KZHffEeWz.",
  "target": {"entity_type": "IfcWall", "guids": []},
  "evaluation": {},
  "tags": ["create", "IfcWall", "scene:R", "direct", "no_anchor"]
}
```

## Loading

```python
from datasets import load_dataset

ds = load_dataset("BIM-Edit/BIM-Edit-Tasks", split="train")
print(len(ds))              # 324
print(ds[0]["prompt"])

walls = ds.filter(lambda r: r["target"]["entity_type"] == "IfcWall")
topological = ds.filter(lambda r: r["category"] == "topological")
```

Or read it directly, since it is a single JSONL file:

```python
import json

with open("tasks.jsonl", encoding="utf-8") as f:
    tasks = [json.loads(line) for line in f]
```