| # Examples |
|
|
| Two kinds of artifact, both verbatim and unedited: |
|
|
| **1. Full-size training rows** — so the scale of the corpus is visible, not just described. |
|
|
| | file | what it is | size | |
| |---|---|---| |
| | `training_example_ru_whole_building_trunk.md` | Russian. One whole building read from the architecture angle — massing, datum stack, apartment programme — synthesized from an IFC extraction. | **42,685 chars** (5,585 scenario / 12,787 reasoning / 25,510 code) | |
| | `training_example_en_rebar_forensics.md` | English. Rebar-to-host integrity forensics on a stadium raker frame: the model refuses the detailer's hypothesis and sets a falsifiable anchor before touching anything. | **17,147 chars** | |
|
|
| Projects are de-identified. In the Russian record, every figure marked `store-точно` ("store-exact") |
| came out of the IFC extraction, not out of the model; where the model reconstructs instead of reads, |
| it says so. |
|
|
| Both rows were compiled against the real Revit assemblies while preparing this release: |
|
|
| | record | builds on | |
| |---|---| |
| | Russian whole-building trunk (standalone file, 25,511 chars of C#) | 2021, 2024, 2026 — verified | |
| | English rebar forensics (3 fragments) | **2024, 2025, 2026 only** — fails 2021–2023 with `CS1503`, because it uses `ElementId.Value`, which exists from 2024 (earlier versions want `IntegerValue`) | |
|
|
| The English row's version lock is not a mistake for a Revit-2026 corpus, but it makes the point of the |
| `wall_with_window` example below: hand-written code carries a silent version dependency, while the |
| compiler emits correct code for all six versions from one source. |
|
|
| **2. A worked KIR example: the same building fragment in KIR and in emitted C#** |
|
|
| A 6 m exterior wall with a window in it, on one level. |
|
|
| | file | what it is | size | |
| |---|---|---| |
| | `wall_with_window.kir.json` | what the **model** writes — the typed IR program | 336 chars / **132 tokens** | |
| | `wall_with_window.emitted.cs` | what the **compiler** produced from it, verbatim | 10,010 chars / **193 lines** / 3,125 tokens | |
|
|
| Token counts measured with this repository's own `tokenizer.json`. Line count: |
| `wc -l wall_with_window.emitted.cs`. |
|
|
| The emitted file is unedited compiler output (`kukai.ir.compiler.compile_program`, target Revit 2026). |
| Read it and note what the model never had to get right: foot/millimetre conversion at every boundary, |
| transaction start-status checking, a rollback on every null return, `FamilySymbol.Activate()` followed |
| by `doc.Regenerate()`, resolving the window's level from its host wall, and a post-`Regenerate()` |
| read-back that compares the built element's actual position against the request to within 10 mm and |
| rolls the whole transaction back if it disagrees. |
|
|
| That is the 23.7× token ratio in the model card, in full, so you can check it rather than trust it. |
|
|
| ## Verification |
|
|
| The emitted file in this folder was compiled against the **real Autodesk Revit assemblies** |
| (`RevitAPI.dll`, `RevitAPIUI.dll`) for every supported version, through the same compile gate the |
| project uses in production: |
|
|
| ``` |
| Revit 2021: OK Revit 2024: OK |
| Revit 2022: OK Revit 2025: OK |
| Revit 2023: OK Revit 2026: OK |
| => 6/6 |
| ``` |
|
|
| So the 193 lines are not illustrative pseudo-code: every API call in them exists and type-checks on |
| all six Revit versions, from one 336-character source program. |
|
|
