Datasets:

AIGeeksGroup
/

SimCode4D

@@ -20,71 +20,22 @@ configs:
 # SimCode4D DEV15
-SimCode4D is the 15-scene **development benchmark** for [Code2Worlds-2](https://github.com/AIGeeksGroup/Code2Worlds-2). It contains three scenes in each of five domains: indoor, industrial, natural outdoor, service, and urban. The Data Studio table is a searchable scene index, not a leaderboard or method result.
-This repository provides the scene prompts, 15 directly browsable MuJoCo runtime worlds and their required assets, 12 archived Blender authoring projects, and 18 Blender animation frames from six scenes. The [benchmark evaluator and metric definitions](https://github.com/AIGeeksGroup/Code2Worlds-2/tree/main/benchmark/simcode4d_unified) are maintained on GitHub.
-## Layout
-| Path | Contents |
-| --- | --- |
-| `data/dev15.jsonl` | One index record per scene: task, runtime model path, optional Blender project, and optional preview paths |
-| `data/runtime/artifacts/<case_id>/` | MuJoCo scene XML for each of the 15 cases |
-| `data/runtime/assets/`, `data/runtime/vendor/` | Meshes, robot descriptions, and notices required by those scenes |
-| `blender_sources/<case_id>/scene.blend` | 12 available authoring projects |
-| `blender_sources/<case_id>/preview_*.png` | Three start/middle/end frames for each of six natural outdoor and service scenes |
-| `THIRD_PARTY_NOTICES.md` | Attribution and license notices for included third-party assets |
-The runtime files are presented individually, without a download archive. Their `artifacts/`, `assets/`, and `vendor/` structure preserves the relative references needed to load the XML worlds.
-| Domain | Scenes | Scene folders with previews |
-| --- | ---: | --- |
-| Indoor | 3 | No published preview |
-| Industrial | 3 | No published preview |
-| Natural outdoor | 3 | [NAT-01-rock-kick](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/tree/main/blender_sources/NAT-01-rock-kick), [NAT-02-hills-turn-stop](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/tree/main/blender_sources/NAT-02-hills-turn-stop), [NAT-03-water-land](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/tree/main/blender_sources/NAT-03-water-land) |
-| Service | 3 | [SER-01-medicine](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/tree/main/blender_sources/SER-01-medicine), [SER-02-airport-luggage](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/tree/main/blender_sources/SER-02-airport-luggage), [SER-03-hotel-drawer](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/tree/main/blender_sources/SER-03-hotel-drawer) |
-| Urban | 3 | No published preview |
-Only the natural outdoor and service scenes have published preview images. These are frames from existing Blender animations. Earlier simplified simulation screenshots for indoor, industrial, and urban scenes were omitted because they did not represent the appearance of the authored worlds well. Missing preview paths are `null` in the index; the runtime scene is still present.
-The original `.blend` sources for **IND-01**, **IND-02**, and **IND-03** were not archived in the available benchmark handover. IND-01 and IND-02 point to the same missing pre-Module-1 source; IND-03 points to a separate missing source. Their compiled MuJoCo worlds are available in `data/runtime/`. No different Blender scene is substituted for them.
-## Download and open a runtime world
-Download the index and its runtime assets without downloading the large Blender projects:
-```bash
-hf download AIGeeksGroup/SimCode4D --repo-type dataset \
-  --include "data/*" --local-dir SimCode4D
-```
-The `runtime_model` field in each JSONL row is a path relative to the downloaded repository. For example:
-```python
-import json
-from pathlib import Path
-import mujoco
-root = Path("SimCode4D")
-rows = [json.loads(line) for line in (root / "data/dev15.jsonl").read_text().splitlines()]
-case = next(row for row in rows if row["case_id"] == "INS-01-agv-aisle")
-model = mujoco.MjModel.from_xml_path(str(root / case["runtime_model"]))
-print(model.ngeom)
-```
-Download one Blender project and its previews when needed:
-```bash
-hf download AIGeeksGroup/SimCode4D --repo-type dataset \
-  --include "blender_sources/INS-01-agv-aisle/*" --local-dir SimCode4D
-```
-Opening Blender projects may require Blender and the original scene-generation dependencies. All 15 runtime models can be loaded from the published `data/runtime/` files.
 ## Evaluation
-Use the [GitHub benchmark guide](https://github.com/AIGeeksGroup/Code2Worlds-2/tree/main/benchmark/simcode4d_unified) to run integrity checks or score a generated world. It defines WorldBuildPass@1, TaskOverlayPass@1, entity binding F1, task success, mechanics, kinematics, contact/friction, safety, and robust task success. This release is a **development/validation** split, not an unseen test set. The reference matrix checks benchmark self-consistency; it is not a Code2Worlds-2 method score.
-## Provenance and use
-The runtime data comes from the public Code2Worlds-2 development suite. Each available Blender file was checked against that suite's source record before upload. Third-party asset notices are in [`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md). No repository-wide license is asserted here; review the notices and applicable project terms before reuse.

 # SimCode4D DEV15
+The 15-scene **development benchmark** for [Code2Worlds-2](https://github.com/AIGeeksGroup/Code2Worlds-2). It has three scenes each in indoor, industrial, natural outdoor, service, and urban domains. The Dataset Viewer lists scene prompts and file paths; it does not show method scores.
+## Data
+- [`data/dev15.jsonl`](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/blob/main/data/dev15.jsonl): the 15-case index and task prompts.
+- [`data/runtime/`](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/tree/main/data/runtime): all 15 loadable MuJoCo worlds and their required assets, available as individual files.
+- [`blender_sources/`](https://huggingface.co/datasets/AIGeeksGroup/SimCode4D/tree/main/blender_sources): 12 archived `.blend` projects. Six natural outdoor and service cases also contain three Blender animation frames beside their project file.
+The three indoor authoring `.blend` files were not archived; their runnable worlds are included. Missing `.blend` and preview paths are `null` in the index. This is a development/validation split, not a hidden test set.
 ## Evaluation
+The [version-matched GitHub benchmark guide](https://github.com/AIGeeksGroup/Code2Worlds-2/blob/ca35f25cda19aadf1d211ae3e55e390905507f1d/benchmark/simcode4d_unified/README.md) provides the frozen case specifications, gold labels, probes, validators, metric implementations, and submission scorer. Its checkout already includes the same runtime assets, so use that guide to run or score the benchmark; no second download of this dataset is needed for scoring.
+The public scorer reports runtime entity-binding F1. The paper's source-name binding F1 requires authoring-source catalogs outside this release and is reported as `null`.
+## Provenance
+See [`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md) for included robot-asset notices. No dataset-wide license is asserted here.