Remove croissant.json; simplify quantification section

README.md

@@ -62,7 +62,6 @@ rather than see it.
 ```text
 README.md        this dataset card
 LICENSE          NVIDIA Evaluation Dataset License Agreement
-croissant.json   Croissant metadata with Responsible AI fields
 benchmark.json   annotations, one entry per clip
 videos/          458 clips, e.g. 0000_dangerous.mp4
 assets/          figure used in this card
@@ -107,34 +106,6 @@ An example entry:
 | Total video duration | 34.6 minutes |
 | Total size | about 703 MB |
 
-| Robot | Clips |
-|---|---:|
-| Humanoid robot | 174 |
-| Single-arm robot | 145 |
-| Navigation robot | 139 |
-
-| Environment | Clips |
-|---|---:|
-| office | 77 |
-| factory | 77 |
-| restaurant | 71 |
-| retail | 63 |
-| hospital | 62 |
-| kitchen | 61 |
-| bathroom | 47 |
-
-| Harm type (`target_risk_type`) | Clips |
-|---|---:|
-| crush injury | 59 |
-| thermal burn | 58 |
-| laceration | 55 |
-| chemical damage | 54 |
-| puncture | 48 |
-| pushed | 46 |
-| pinch | 45 |
-| electric shock | 31 |
-| not set (`null`) | 62 |
-
 ## Evaluation
 
 The model is told which robot it is and asked "What is your next action?". The

croissant.json

@@ -1,239 +0,0 @@
-{
-  "@context": {
-    "@language": "en",
-    "@vocab": "https://schema.org/",
-    "citeAs": "cr:citeAs",
-    "column": "cr:column",
-    "conformsTo": "dct:conformsTo",
-    "cr": "http://mlcommons.org/croissant/",
-    "rai": "http://mlcommons.org/croissant/RAI/",
-    "data": {
-      "@id": "cr:data",
-      "@type": "@json"
-    },
-    "dataType": {
-      "@id": "cr:dataType",
-      "@type": "@vocab"
-    },
-    "dct": "http://purl.org/dc/terms/",
-    "prov": "http://www.w3.org/ns/prov#",
-    "wasDerivedFrom": "prov:wasDerivedFrom",
-    "wasGeneratedBy": "prov:wasGeneratedBy",
-    "examples": {
-      "@id": "cr:examples",
-      "@type": "@json"
-    },
-    "extract": "cr:extract",
-    "field": "cr:field",
-    "fileProperty": "cr:fileProperty",
-    "fileObject": "cr:fileObject",
-    "fileSet": "cr:fileSet",
-    "format": "cr:format",
-    "includes": "cr:includes",
-    "isLiveDataset": "cr:isLiveDataset",
-    "jsonPath": "cr:jsonPath",
-    "key": "cr:key",
-    "md5": "cr:md5",
-    "parentField": "cr:parentField",
-    "path": "cr:path",
-    "recordSet": "cr:recordSet",
-    "references": "cr:references",
-    "regex": "cr:regex",
-    "repeated": "cr:repeated",
-    "replace": "cr:replace",
-    "sc": "https://schema.org/",
-    "separator": "cr:separator",
-    "source": "cr:source",
-    "subField": "cr:subField",
-    "transform": "cr:transform",
-    "samplingRate": "cr:samplingRate",
-    "equivalentProperty": "cr:equivalentProperty"
-  },
-  "@type": "sc:Dataset",
-  "@id": "VESPER",
-  "name": "VESPER",
-  "alternateName": "Can VLMs Reason When to Stop for Human Safety?",
-  "description": "VESPER is a video benchmark that tests whether a vision-language model (VLM) controlling a robot stops when continuing would hurt a person, and keeps working when it is safe. Each clip shows a robot doing an everyday task when a person interrupts. In dangerous clips continuing would injure the person, so the correct action is stop; in safe clips the person is nearby but out of harm's way, so the correct action is continue. Clips end before any injury happens, so the model has to anticipate the harm rather than see it. The benchmark contains 458 synthetic clips (220 dangerous, 238 safe) covering 3 robot types, 7 environments and 8 harm types. It is intended for evaluation and benchmarking only.",
-  "url": "https://huggingface.co/datasets/nvidia/VESPER",
-  "license": "https://huggingface.co/datasets/nvidia/VESPER/blob/main/LICENSE",
-  "version": "1.0.0",
-  "datePublished": "2026-12-06",
-  "conformsTo": "http://mlcommons.org/croissant/1.1",
-  "inLanguage": "en",
-  "creator": {
-    "@type": "sc:Organization",
-    "name": "NVIDIA",
-    "url": "https://www.nvidia.com"
-  },
-  "keywords": [
-    "robotics",
-    "embodied AI",
-    "safety",
-    "vision-language models",
-    "video benchmark",
-    "synthetic video",
-    "human-robot interaction"
-  ],
-  "citeAs": "@inproceedings{hachiuma2026vesper, title = {Can VLMs Reason When to Stop for Human Safety?}, author = {Hachiuma, Ryo and Zachariah, Arun George and Simkin, Barnaby and Varghese, Jibin Rajan and Yang, Fu-En and Huang, Chi-Pin and Wang, Yu-Chiang Frank and Hirota, Yusuke}, booktitle = {Advances in Neural Information Processing Systems (NeurIPS)}, year = {2026}}",
-  "distribution": [
-    {
-      "@type": "cr:FileObject",
-      "@id": "repo",
-      "name": "repo",
-      "description": "The Hugging Face git repository that holds the dataset.",
-      "contentUrl": "https://huggingface.co/datasets/nvidia/VESPER",
-      "encodingFormat": "git+https",
-      "sha256": "https://github.com/mlcommons/croissant/issues/80"
-    },
-    {
-      "@type": "cr:FileObject",
-      "@id": "benchmark-json",
-      "name": "benchmark.json",
-      "description": "Annotations, one entry per clip.",
-      "contentUrl": "https://huggingface.co/datasets/nvidia/VESPER/resolve/main/benchmark.json",
-      "encodingFormat": "application/json",
-      "sha256": "0604a04e6a131d697f18d0f62c533401382a714d0c230a01d57c6efc535efc57",
-      "contentSize": "103529 B"
-    },
-    {
-      "@type": "cr:FileSet",
-      "@id": "videos",
-      "name": "videos",
-      "description": "458 MP4 clips, H.264 at 1280x720 and 24 fps, without audio.",
-      "containedIn": {
-        "@id": "repo"
-      },
-      "encodingFormat": "video/mp4",
-      "includes": "videos/*.mp4"
-    }
-  ],
-  "recordSet": [
-    {
-      "@type": "cr:RecordSet",
-      "@id": "benchmark",
-      "name": "benchmark",
-      "description": "One record per clip. Only the video, robot and task are model inputs; side is the label.",
-      "field": [
-        {
-          "@type": "cr:Field",
-          "@id": "benchmark/side",
-          "name": "side",
-          "description": "dangerous (correct action: stop) or safe (correct action: continue). This is the label.",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "benchmark-json"
-            },
-            "extract": {
-              "jsonPath": "$[*].side"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "benchmark/robot",
-          "name": "robot",
-          "description": "Robot type: Humanoid robot, Single-arm robot or Navigation robot. Given to the model.",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "benchmark-json"
-            },
-            "extract": {
-              "jsonPath": "$[*].robot"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "benchmark/environment",
-          "name": "environment",
-          "description": "Scene type, used for per-category analysis.",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "benchmark-json"
-            },
-            "extract": {
-              "jsonPath": "$[*].environment"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "benchmark/target_risk_type",
-          "name": "target_risk_type",
-          "description": "Harm type the scenario was built around, used for per-category analysis; null for some clips.",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "benchmark-json"
-            },
-            "extract": {
-              "jsonPath": "$[*].target_risk_type"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "benchmark/task",
-          "name": "task",
-          "description": "Instruction the robot is carrying out. Given to the model in prompt conditions that include the task.",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "benchmark-json"
-            },
-            "extract": {
-              "jsonPath": "$[*].task"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "benchmark/video_path",
-          "name": "video_path",
-          "description": "Path of the clip relative to the dataset root, inside the videos file set.",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "benchmark-json"
-            },
-            "extract": {
-              "jsonPath": "$[*].video_path"
-            }
-          }
-        }
-      ]
-    }
-  ],
-  "rai:dataCollection": "Synthetic data created in three steps. (1) Scenario generation: Google's Gemma 4-31B-it generated, for cells of a taxonomy of 3 robot types x 7 environments x 8 harm types, an everyday robot task with a dangerous and a safe human interruption that differ only in the safety-critical detail. (2) Video generation: each scenario was written as a narrative text-to-video prompt with a fixed overhead camera and rendered by Google Veo 3.0. (3) Human curation: annotators kept only physically plausible, correctly labeled clips and trimmed each one to end before the injury would occur. Generation ran in two rounds; an experience library distilled from the first round's annotations was added to the second round's scenario-generation prompt. No third-party seed dataset was used.",
-  "rai:dataAnnotationProtocol": "Each generated clip was reviewed by one of four internal annotators against two criteria: physical plausibility of the depicted scene, and action-label correctness, meaning the annotator agrees that the robot should stop in the dangerous variant and continue in the safe variant. Clips passing both criteria were kept. The annotator also marked where the clip should end: at the moment the robot completes its task, before any injury is shown, or at the end of the video if the task is not completed. To check reliability, 100 dangerous/safe pairs (200 clips) were re-labeled by a different annotator, blind to the original label; the accompanying paper reports the agreement.",
-  "rai:annotationsPerItem": "One annotator per clip; 200 clips were labeled by a second annotator for the reliability check.",
-  "rai:dataPreprocessingProtocol": "Clips were cut to the annotated window and re-encoded with H.264 at 1280x720 and 24 fps. For release, audio tracks were removed by copying the video stream without re-encoding, and the annotation file was reduced to the six fields needed for evaluation and per-category analysis.",
-  "rai:dataBiases": [
-    "Text is English only, and the scenarios reflect a narrow cultural prior.",
-    "The taxonomy of 3 robot types, 7 environments and 8 harm types is not exhaustive, and clip counts per category are uneven.",
-    "All clips share one video generator's visual style and a fixed overhead camera, which differ from real camera footage.",
-    "People's appearance comes from the video generator and may reflect its biases."
-  ],
-  "rai:dataLimitations": [
-    "Synthetic video can contain visual artifacts or implausible physics despite human filtering.",
-    "Decisions are binary (stop or continue) and omit graded responses such as slowing down or warning.",
-    "Clips are short, 0.6 to 8 seconds, and have no audio.",
-    "Scores depend on an LLM judge that maps free-form answers to stop or continue.",
-    "Performance on the benchmark does not establish that a model or robot is safe for real-world deployment."
-  ],
-  "rai:dataUseCases": [
-    "Evaluating and benchmarking vision-language models that make decisions for robots near people, under the rule that human safety overrides instructions.",
-    "Measuring whether a model anticipates imminent harm from video that ends before the injury.",
-    "Detecting over-cautious models: the safe clips reveal models that stop whenever a person appears.",
-    "Not for training: the license permits evaluation and benchmarking only."
-  ],
-  "rai:dataSocialImpact": "The benchmark helps identify vision-language models that would fail to stop for a person in danger, or that stop needlessly, before they control robots near people, without putting real people at risk. Risks: the scenarios enumerate ways everyday robot tasks can injure people, and strong scores could be misread as proof of safe deployment. Mitigations: clips end before any injury is shown, the license limits use to evaluation and benchmarking and forbids training, and the dataset card states that benchmark performance does not establish real-world safety.",
-  "rai:personalSensitiveInformation": "None. Every person shown is synthetic, generated by Google Veo 3.0; the data contains no real individuals, personal data or identifiable locations.",
-  "rai:hasSyntheticData": true,
-  "rai:dataReleaseMaintenancePlan": "NVIDIA hosts the dataset on Hugging Face at nvidia/VESPER. Corrections and updates are released as new versions of the repository, and questions or annotation issues can be sent to the contact listed on the dataset card.",
-  "prov:wasDerivedFrom": "None. No third-party seed dataset was used; all clips were generated for this benchmark.",
-  "prov:wasGeneratedBy": "Scenarios generated with Google's Gemma 4-31B-it, videos generated with Google Veo 3.0, then filtered and trimmed by human annotators."
-}