Upload croissant.json with huggingface_hub

croissant.json

@@ -1,283 +1,302 @@
 {
-  "@context": {
-    "@language": "en",
-    "@vocab": "https://schema.org/",
-    "sc": "https://schema.org/",
-    "cr": "http://mlcommons.org/croissant/",
-    "rai": "http://mlcommons.org/croissant/RAI/",
-    "dct": "http://purl.org/dc/terms/"
-  },
-  "@type": "sc:Dataset",
-  "dct:conformsTo": "http://mlcommons.org/croissant/1.0",
-  "name": "stem-gym-benchmark",
-  "description": "STEMGym is a Gymnasium-based benchmark for evaluating autonomous dose-efficient scanning transmission electron microscopy (STEM) agents. The dataset provides 15 simulated STEM specimens (5 materials x 3 difficulty levels) as HDF5 world files, each containing high-resolution HAADF-STEM tile grids, overview images, and ground-truth annotations (atom positions, defect types, phase maps). It also includes pre-trained analyst model checkpoints and RL navigation baselines.",
-  "url": "https://huggingface.co/datasets/anonneuripsuser/stem-gym-benchmark",
-  "license": "https://creativecommons.org/licenses/by/4.0/",
-  "version": "1.0.0",
-  "datePublished": "2026-04-02",
-  "dateCreated": "2025-12-01",
-  "dateModified": "2026-04-02",
-  "inLanguage": "en",
-  "isLiveDataset": false,
-  "keywords": [
-    "stem-microscopy",
-    "electron-microscopy",
-    "materials-science",
-    "gymnasium",
-    "reinforcement-learning",
-    "benchmark",
-    "dose-efficiency",
-    "autonomous-microscopy",
-    "simulation",
-    "HDF5"
-  ],
-  "creator": {
-    "@type": "sc:Person",
-    "name": "Anonymous"
-  },
-  "publisher": {
-    "@type": "sc:Organization",
-    "name": "Anonymous"
-  },
-  "citeAs": "@inproceedings{anonymous2026stemgym, title={STEMGym: A Gymnasium Environment for Benchmarking Dose-Efficient Autonomous Scanning Transmission Electron Microscopy}, author={Anonymous}, year={2026}}",
-
-  "distribution": [
-    {
-      "@type": "cr:FileSet",
-      "@id": "world-files",
-      "name": "world-files",
-      "description": "Simulated STEM specimen worlds as HDF5 files. 5 materials (SrTiO3, BaTiO3, SiGe, GaN, Pt nanoparticles) at 3 difficulty levels (easy, medium, hard), plus test and replay worlds.",
-      "containedIn": {"@id": "hf-repo"},
-      "encodingFormat": "application/x-hdf5",
-      "includes": "worlds/*.h5"
+    "@context": {
+        "@language": "en",
+        "@vocab": "https://schema.org/",
+        "sc": "https://schema.org/",
+        "cr": "http://mlcommons.org/croissant/",
+        "rai": "http://mlcommons.org/croissant/RAI/",
+        "dct": "http://purl.org/dc/terms/"
     },
-    {
-      "@type": "cr:FileSet",
-      "@id": "main-checkpoints",
-      "name": "main-checkpoints",
-      "description": "Pre-trained analyst model checkpoints (AtomFinderUNet, DefectClassifierCNN, PhaseIdentifierResNet) and RL navigation baselines (DQN, PPO, SAC).",
-      "containedIn": {"@id": "hf-repo"},
-      "encodingFormat": "application/octet-stream",
-      "includes": "checkpoints/*.pt"
+    "@type": "sc:Dataset",
+    "dct:conformsTo": "http://mlcommons.org/croissant/1.0",
+    "name": "stem-gym-benchmark",
+    "description": "STEMGym is a Gymnasium-based benchmark for evaluating autonomous dose-efficient scanning transmission electron microscopy (STEM) agents. The dataset provides 15 simulated STEM specimens (5 materials x 3 difficulty levels) as HDF5 world files, each containing high-resolution HAADF-STEM tile grids, overview images, and ground-truth annotations (atom positions, defect types, phase maps). It also includes pre-trained analyst model checkpoints and RL navigation baselines.",
+    "url": "https://huggingface.co/datasets/anonneuripsuser/stem-gym-benchmark",
+    "license": "https://creativecommons.org/licenses/by/4.0/",
+    "version": "1.0.0",
+    "datePublished": "2026-04-02",
+    "dateCreated": "2025-12-01",
+    "dateModified": "2026-04-02",
+    "inLanguage": "en",
+    "isLiveDataset": false,
+    "keywords": [
+        "stem-microscopy",
+        "electron-microscopy",
+        "materials-science",
+        "gymnasium",
+        "reinforcement-learning",
+        "benchmark",
+        "dose-efficiency",
+        "autonomous-microscopy",
+        "simulation",
+        "HDF5"
+    ],
+    "creator": {
+        "@type": "sc:Person",
+        "name": "Anonymous"
     },
-    {
-      "@type": "cr:FileSet",
-      "@id": "rl-checkpoints",
-      "name": "rl-checkpoints",
-      "description": "Stable-Baselines3 RL agent checkpoints (DQN, PPO, SAC) for navigation baselines.",
-      "containedIn": {"@id": "hf-repo"},
-      "encodingFormat": "application/zip",
-      "includes": "checkpoints/*.zip"
+    "publisher": {
+        "@type": "sc:Organization",
+        "name": "Anonymous"
     },
-    {
-      "@type": "cr:FileSet",
-      "@id": "transfer-checkpoints",
-      "name": "transfer-checkpoints",
-      "description": "Material-specific analyst models trained on individual materials for transfer learning experiments. One set per material (SrTiO3, BaTiO3, SiGe, GaN).",
-      "containedIn": {"@id": "hf-repo"},
-      "encodingFormat": "application/octet-stream",
-      "includes": "checkpoints/transfer/*/*.pt"
-    },
-    {
-      "@type": "cr:FileObject",
-      "@id": "hf-repo",
-      "name": "hf-repo",
-      "description": "HuggingFace dataset repository.",
-      "contentUrl": "https://huggingface.co/datasets/anonneuripsuser/stem-gym-benchmark/tree/main",
-      "encodingFormat": "git+https",
-      "sha256": ""
-    }
-  ],
-
-  "recordSet": [
-    {
-      "@type": "cr:RecordSet",
-      "@id": "world-metadata",
-      "name": "world-metadata",
-      "description": "Per-world metadata stored in HDF5 /metadata/ group.",
-      "field": [
-        {
-          "@type": "cr:Field",
-          "@id": "world-metadata/filename",
-          "name": "filename",
-          "description": "World HDF5 filename.",
-          "dataType": "sc:Text",
-          "source": {
-            "fileSet": {"@id": "world-files"},
-            "extract": {"fileProperty": "filename"}
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "world-metadata/pixel_size_nm",
-          "name": "pixel_size_nm",
-          "description": "Physical pixel size in nanometers.",
-          "dataType": "sc:Float"
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "world-metadata/tile_size_px",
-          "name": "tile_size_px",
-          "description": "Tile dimensions in pixels (128).",
-          "dataType": "sc:Integer"
-        },
+    "cr:citeAs": "@inproceedings{anonymous2026stemgym, title={STEMGym: A Gymnasium Environment for Benchmarking Dose-Efficient Autonomous Scanning Transmission Electron Microscopy}, author={Anonymous}, year={2026}}",
+    "distribution": [
         {
-          "@type": "cr:Field",
-          "@id": "world-metadata/grid_shape",
-          "name": "grid_shape",
-          "description": "Grid dimensions as (rows, cols).",
-          "dataType": "sc:Text"
+            "@type": "cr:FileSet",
+            "@id": "world-files",
+            "name": "world-files",
+            "description": "Simulated STEM specimen worlds as HDF5 files. 5 materials (SrTiO3, BaTiO3, SiGe, GaN, Pt nanoparticles) at 3 difficulty levels (easy, medium, hard), plus test and replay worlds.",
+            "containedIn": {
+                "@id": "hf-repo"
+            },
+            "encodingFormat": "application/x-hdf5",
+            "cr:includes": "worlds/*.h5"
         },
         {
-          "@type": "cr:Field",
-          "@id": "world-metadata/fov_nm",
-          "name": "fov_nm",
-          "description": "Field of view in nanometers.",
-          "dataType": "sc:Float"
+            "@type": "cr:FileSet",
+            "@id": "main-checkpoints",
+            "name": "main-checkpoints",
+            "description": "Pre-trained analyst model checkpoints (AtomFinderUNet, DefectClassifierCNN, PhaseIdentifierResNet) and RL navigation baselines (DQN, PPO, SAC).",
+            "containedIn": {
+                "@id": "hf-repo"
+            },
+            "encodingFormat": "application/octet-stream",
+            "cr:includes": "checkpoints/*.pt"
         },
         {
-          "@type": "cr:Field",
-          "@id": "world-metadata/scenario",
-          "name": "scenario",
-          "description": "Material/scenario identifier.",
-          "dataType": "sc:Text"
+            "@type": "cr:FileSet",
+            "@id": "rl-checkpoints",
+            "name": "rl-checkpoints",
+            "description": "Stable-Baselines3 RL agent checkpoints (DQN, PPO, SAC) for navigation baselines.",
+            "containedIn": {
+                "@id": "hf-repo"
+            },
+            "encodingFormat": "application/zip",
+            "cr:includes": "checkpoints/*.zip"
         },
         {
-          "@type": "cr:Field",
-          "@id": "world-metadata/difficulty",
-          "name": "difficulty",
-          "description": "Difficulty level: easy, medium, or hard.",
-          "dataType": "sc:Text"
-        }
-      ]
-    },
-    {
-      "@type": "cr:RecordSet",
-      "@id": "world-images",
-      "name": "world-images",
-      "description": "STEM image data stored in each HDF5 world file.",
-      "field": [
-        {
-          "@type": "cr:Field",
-          "@id": "world-images/overview",
-          "name": "overview",
-          "description": "Low-magnification overview image of the full specimen, stored at HDF5 path /overview. Shape: (H, W) float32.",
-          "dataType": "sc:ImageObject"
+            "@type": "cr:FileSet",
+            "@id": "transfer-checkpoints",
+            "name": "transfer-checkpoints",
+            "description": "Material-specific analyst models trained on individual materials for transfer learning experiments. One set per material (SrTiO3, BaTiO3, SiGe, GaN).",
+            "containedIn": {
+                "@id": "hf-repo"
+            },
+            "encodingFormat": "application/octet-stream",
+            "cr:includes": "checkpoints/transfer/*/*.pt"
         },
         {
-          "@type": "cr:Field",
-          "@id": "world-images/tiles",
-          "name": "tiles",
-          "description": "High-resolution STEM-HAADF tile images stored at HDF5 path /tiles/{row}_{col}. Each tile is (128, 128) float32 normalized. Tiles arranged in 8x8 grid with 4px overlap (stride=124).",
-          "dataType": "sc:ImageObject",
-          "repeated": true
+            "@type": "cr:FileObject",
+            "@id": "hf-repo",
+            "name": "hf-repo",
+            "description": "HuggingFace dataset repository.",
+            "contentUrl": "https://huggingface.co/datasets/anonneuripsuser/stem-gym-benchmark/tree/main",
+            "encodingFormat": "git+https",
+            "sha256": ""
         }
-      ]
-    },
-    {
-      "@type": "cr:RecordSet",
-      "@id": "ground-truth",
-      "name": "ground-truth",
-      "description": "Ground-truth annotations stored in each HDF5 world file at /ground_truth/.",
-      "field": [
-        {
-          "@type": "cr:Field",
-          "@id": "ground-truth/atom_positions",
-          "name": "atom_positions",
-          "description": "Atomic column positions in nanometers. Shape: (N, 2) float64.",
-          "dataType": "sc:Float",
-          "repeated": true
-        },
+    ],
+    "recordSet": [
         {
-          "@type": "cr:Field",
-          "@id": "ground-truth/atom_types",
-          "name": "atom_types",
-          "description": "Atom type labels: 0=pristine, 1=vacancy, 2=substitution. Shape: (N,) int.",
-          "dataType": "sc:Integer",
-          "repeated": true
+            "@type": "cr:RecordSet",
+            "@id": "world-metadata",
+            "name": "world-metadata",
+            "description": "Per-world metadata stored in HDF5 /metadata/ group.",
+            "field": [
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-metadata/filename",
+                    "name": "filename",
+                    "description": "World HDF5 filename.",
+                    "dataType": "sc:Text",
+                    "source": {
+                        "fileSet": {
+                            "@id": "world-files"
+                        },
+                        "extract": {
+                            "fileProperty": "filename"
+                        }
+                    }
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-metadata/pixel_size_nm",
+                    "name": "pixel_size_nm",
+                    "description": "Physical pixel size in nanometers.",
+                    "dataType": "sc:Float"
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-metadata/tile_size_px",
+                    "name": "tile_size_px",
+                    "description": "Tile dimensions in pixels (128).",
+                    "dataType": "sc:Integer"
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-metadata/grid_shape",
+                    "name": "grid_shape",
+                    "description": "Grid dimensions as (rows, cols).",
+                    "dataType": "sc:Text"
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-metadata/fov_nm",
+                    "name": "fov_nm",
+                    "description": "Field of view in nanometers.",
+                    "dataType": "sc:Float"
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-metadata/scenario",
+                    "name": "scenario",
+                    "description": "Material/scenario identifier.",
+                    "dataType": "sc:Text"
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-metadata/difficulty",
+                    "name": "difficulty",
+                    "description": "Difficulty level: easy, medium, or hard.",
+                    "dataType": "sc:Text"
+                }
+            ]
         },
         {
-          "@type": "cr:Field",
-          "@id": "ground-truth/defect_mask",
-          "name": "defect_mask",
-          "description": "Boolean mask indicating defective atoms. Shape: (N,) bool.",
-          "dataType": "sc:Boolean",
-          "repeated": true
+            "@type": "cr:RecordSet",
+            "@id": "world-images",
+            "name": "world-images",
+            "description": "STEM image data stored in each HDF5 world file.",
+            "field": [
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-images/overview",
+                    "name": "overview",
+                    "description": "Low-magnification overview image of the full specimen, stored at HDF5 path /overview. Shape: (H, W) float32.",
+                    "dataType": "sc:ImageObject"
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "world-images/tiles",
+                    "name": "tiles",
+                    "description": "High-resolution STEM-HAADF tile images stored at HDF5 path /tiles/{row}_{col}. Each tile is (128, 128) float32 normalized. Tiles arranged in 8x8 grid with 4px overlap (stride=124).",
+                    "dataType": "sc:ImageObject",
+                    "repeated": true
+                }
+            ]
         },
         {
-          "@type": "cr:Field",
-          "@id": "ground-truth/defect_types",
-          "name": "defect_types",
-          "description": "Defect type strings: pristine, vacancy, or substitution. Shape: (N,) str.",
-          "dataType": "sc:Text",
-          "repeated": true
+            "@type": "cr:RecordSet",
+            "@id": "ground-truth",
+            "name": "ground-truth",
+            "description": "Ground-truth annotations stored in each HDF5 world file at /ground_truth/.",
+            "field": [
+                {
+                    "@type": "cr:Field",
+                    "@id": "ground-truth/atom_positions",
+                    "name": "atom_positions",
+                    "description": "Atomic column positions in nanometers. Shape: (N, 2) float64.",
+                    "dataType": "sc:Float",
+                    "repeated": true
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "ground-truth/atom_types",
+                    "name": "atom_types",
+                    "description": "Atom type labels: 0=pristine, 1=vacancy, 2=substitution. Shape: (N,) int.",
+                    "dataType": "sc:Integer",
+                    "repeated": true
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "ground-truth/defect_mask",
+                    "name": "defect_mask",
+                    "description": "Boolean mask indicating defective atoms. Shape: (N,) bool.",
+                    "dataType": "sc:Boolean",
+                    "repeated": true
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "ground-truth/defect_types",
+                    "name": "defect_types",
+                    "description": "Defect type strings: pristine, vacancy, or substitution. Shape: (N,) str.",
+                    "dataType": "sc:Text",
+                    "repeated": true
+                },
+                {
+                    "@type": "cr:Field",
+                    "@id": "ground-truth/phase_map",
+                    "name": "phase_map",
+                    "description": "Phase map as integer labels. Shape: (H, W) int32. Optional, present for multi-phase materials (BaTiO3).",
+                    "dataType": "cr:SegmentationMask"
+                }
+            ]
         },
         {
-          "@type": "cr:Field",
-          "@id": "ground-truth/phase_map",
-          "name": "phase_map",
-          "description": "Phase map as integer labels. Shape: (H, W) int32. Optional, present for multi-phase materials (BaTiO3).",
-          "dataType": "cr:SegmentationMask"
-        }
-      ]
-    },
-    {
-      "@type": "cr:RecordSet",
-      "@id": "valid-region",
-      "name": "valid-region",
-      "description": "Boolean mask at /valid_region indicating which pixels contain valid specimen data. Shape: (H, W) bool.",
-      "field": [
-        {
-          "@type": "cr:Field",
-          "@id": "valid-region/mask",
-          "name": "mask",
-          "description": "Per-pixel validity mask for the specimen grid.",
-          "dataType": "sc:Boolean",
-          "repeated": true
+            "@type": "cr:RecordSet",
+            "@id": "valid-region",
+            "name": "valid-region",
+            "description": "Boolean mask at /valid_region indicating which pixels contain valid specimen data. Shape: (H, W) bool.",
+            "field": [
+                {
+                    "@type": "cr:Field",
+                    "@id": "valid-region/mask",
+                    "name": "mask",
+                    "description": "Per-pixel validity mask for the specimen grid.",
+                    "dataType": "sc:Boolean",
+                    "repeated": true
+                }
+            ]
         }
-      ]
-    }
-  ],
-
-  "rai:dataCollection": "All specimen data is computationally simulated. Crystal structures were built with pymatgen from Materials Project entries, defects were introduced programmatically (vacancies, substitutions, phase boundaries, nanoparticles), and STEM-HAADF images were simulated using the abTEM multislice code with frozen-phonon approximation. No real experimental data or human subjects are involved.",
-  "rai:dataCollectionType": ["Experiments", "Software Collection"],
-  "rai:dataCollectionRawData": "Raw simulation outputs are multislice STEM-HAADF images computed by abTEM. These are post-processed into 128x128 tiles with Poisson noise applied at material-specific dose levels (1e3-1e4 e-/A^2). Ground truth atom positions and defect labels are derived directly from the input crystal structures.",
-  "rai:dataCollectionTimeframe": "2025-06-01/2026-01-15",
-  "rai:dataCollectionMissingData": "No missing data. All 15 material-difficulty combinations are complete. Each world contains a full 8x8 tile grid and exhaustive ground-truth annotations for all atoms in the field of view.",
-  "rai:dataPreprocessingProtocol": [
-    "STEM-HAADF images simulated via abTEM multislice with frozen-phonon thermal diffuse scattering.",
-    "Images normalized to [0, 1] range per tile.",
-    "Poisson noise applied at dose levels corresponding to difficulty (easy=1e4, medium=5e3, hard=1e3 e-/A^2).",
-    "Tiles extracted on 124-pixel stride with 4-pixel overlap from full simulated field of view."
-  ],
-  "rai:dataAnnotationProtocol": "Ground truth is deterministic: atom positions, types, and defect labels are derived directly from the input crystal structures used for simulation. Phase maps are computed from the known crystal phase of each unit cell. No manual annotation was performed.",
-  "rai:dataAnnotationPlatform": ["Automated (simulation-derived)"],
-  "rai:dataAnnotationAnalysis": ["Ground truth is exact by construction since labels are derived from the simulation input, not from the output images."],
-  "rai:annotationsPerItem": "Every atom in each world's field of view is annotated with position (nm), type (pristine/vacancy/substitution), and defect status. Phase maps cover every pixel.",
-  "rai:machineAnnotationTools": ["abTEM (multislice STEM simulation)", "pymatgen (crystal structure generation)", "ASE (atomic simulation environment)"],
-  "rai:personalSensitiveInformation": ["Not applicable. This dataset contains only simulated materials science images with no personal or sensitive information."],
-  "rai:dataSocialImpact": "This benchmark advances autonomous electron microscopy, potentially accelerating materials discovery for energy, electronics, and structural applications. No negative social impacts are anticipated from simulated microscopy data.",
-  "rai:dataBiases": [
-    "Simulation bias: abTEM frozen-phonon approximation may not capture all real-world imaging artifacts (drift, contamination, detector nonlinearity).",
-    "Material selection bias: only 5 material systems are represented, all crystalline or nanoparticulate. Amorphous, biological, and polymeric materials are not covered.",
-    "Defect distribution bias: defect concentrations and spatial distributions are synthetically designed and may not match real-world statistical distributions."
-  ],
-  "rai:dataLimitations": [
-    "Simulated images lack experimental artifacts such as sample drift, beam damage accumulation, hydrocarbon contamination, and detector noise beyond Poisson statistics.",
-    "Fixed field of view per material; does not test multi-scale navigation across orders of magnitude.",
-    "Ground truth assumes perfect knowledge of the atomic structure, which is unavailable in real experiments.",
-    "Only HAADF-STEM imaging mode is simulated; other modalities (ABF, DPC, EELS, EDS) are not included."
-  ],
-  "rai:dataUseCases": [
-    "Benchmarking autonomous STEM microscopy agents on dose-efficient specimen characterization.",
-    "Evaluating navigation, perception, and analysis strategies for automated microscopy.",
-    "Training and testing reinforcement learning agents for scientific instrument control.",
-    "Comparing human and AI performance on microscopy tasks via the included human baseline interface."
-  ],
-  "rai:dataReleaseMaintenancePlan": [
-    "Dataset is versioned and archived. New material systems or difficulty levels may be added in future versions.",
-    "Bug fixes and corrections will be released as point versions with changelogs."
-  ],
-  "rai:dataImputationProtocol": "No imputation is performed. All data values are computed deterministically from simulation.",
-  "rai:dataManipulationProtocol": "No post-hoc manipulation. Tile images and ground truth are direct outputs of the simulation pipeline."
-}
+    ],
+    "rai:dataCollection": "All specimen data is computationally simulated. Crystal structures were built with pymatgen from Materials Project entries, defects were introduced programmatically (vacancies, substitutions, phase boundaries, nanoparticles), and STEM-HAADF images were simulated using the abTEM multislice code with frozen-phonon approximation. No real experimental data or human subjects are involved.",
+    "rai:dataCollectionType": [
+        "Experiments",
+        "Software Collection"
+    ],
+    "rai:dataCollectionRawData": "Raw simulation outputs are multislice STEM-HAADF images computed by abTEM. These are post-processed into 128x128 tiles with Poisson noise applied at material-specific dose levels (1e3-1e4 e-/A^2). Ground truth atom positions and defect labels are derived directly from the input crystal structures.",
+    "rai:dataCollectionTimeframe": "2025-06-01/2026-01-15",
+    "rai:dataCollectionMissingData": "No missing data. All 15 material-difficulty combinations are complete. Each world contains a full 8x8 tile grid and exhaustive ground-truth annotations for all atoms in the field of view.",
+    "rai:dataPreprocessingProtocol": [
+        "STEM-HAADF images simulated via abTEM multislice with frozen-phonon thermal diffuse scattering.",
+        "Images normalized to [0, 1] range per tile.",
+        "Poisson noise applied at dose levels corresponding to difficulty (easy=1e4, medium=5e3, hard=1e3 e-/A^2).",
+        "Tiles extracted on 124-pixel stride with 4-pixel overlap from full simulated field of view."
+    ],
+    "rai:dataAnnotationProtocol": "Ground truth is deterministic: atom positions, types, and defect labels are derived directly from the input crystal structures used for simulation. Phase maps are computed from the known crystal phase of each unit cell. No manual annotation was performed.",
+    "rai:dataAnnotationPlatform": [
+        "Automated (simulation-derived)"
+    ],
+    "rai:dataAnnotationAnalysis": [
+        "Ground truth is exact by construction since labels are derived from the simulation input, not from the output images."
+    ],
+    "rai:annotationsPerItem": "Every atom in each world's field of view is annotated with position (nm), type (pristine/vacancy/substitution), and defect status. Phase maps cover every pixel.",
+    "rai:machineAnnotationTools": [
+        "abTEM (multislice STEM simulation)",
+        "pymatgen (crystal structure generation)",
+        "ASE (atomic simulation environment)"
+    ],
+    "rai:personalSensitiveInformation": [
+        "Not applicable. This dataset contains only simulated materials science images with no personal or sensitive information."
+    ],
+    "rai:dataSocialImpact": "This benchmark advances autonomous electron microscopy, potentially accelerating materials discovery for energy, electronics, and structural applications. No negative social impacts are anticipated from simulated microscopy data.",
+    "rai:dataBiases": [
+        "Simulation bias: abTEM frozen-phonon approximation may not capture all real-world imaging artifacts (drift, contamination, detector nonlinearity).",
+        "Material selection bias: only 5 material systems are represented, all crystalline or nanoparticulate. Amorphous, biological, and polymeric materials are not covered.",
+        "Defect distribution bias: defect concentrations and spatial distributions are synthetically designed and may not match real-world statistical distributions."
+    ],
+    "rai:dataLimitations": [
+        "Simulated images lack experimental artifacts such as sample drift, beam damage accumulation, hydrocarbon contamination, and detector noise beyond Poisson statistics.",
+        "Fixed field of view per material; does not test multi-scale navigation across orders of magnitude.",
+        "Ground truth assumes perfect knowledge of the atomic structure, which is unavailable in real experiments.",
+        "Only HAADF-STEM imaging mode is simulated; other modalities (ABF, DPC, EELS, EDS) are not included."
+    ],
+    "rai:dataUseCases": [
+        "Benchmarking autonomous STEM microscopy agents on dose-efficient specimen characterization.",
+        "Evaluating navigation, perception, and analysis strategies for automated microscopy.",
+        "Training and testing reinforcement learning agents for scientific instrument control.",
+        "Comparing human and AI performance on microscopy tasks via the included human baseline interface."
+    ],
+    "rai:dataReleaseMaintenancePlan": "Dataset is versioned and archived. New material systems or difficulty levels may be added in future versions. Bug fixes and corrections will be released as point versions with changelogs.",
+    "rai:dataImputationProtocol": "No imputation is performed. All data values are computed deterministically from simulation.",
+    "rai:dataManipulationProtocol": "No post-hoc manipulation. Tile images and ground truth are direct outputs of the simulation pipeline."
+}