Use rai: prefix for top-level RAI fields (validator literal-key match)

croissant.json

@@ -12,8 +12,6 @@
       "@id": "cr:data",
       "@type": "@json"
     },
-    "dataBiases": "rai:dataBiases",
-    "dataCollection": "rai:dataCollection",
     "dataType": {
       "@id": "cr:dataType",
       "@type": "@vocab"
@@ -33,7 +31,6 @@
     "md5": "cr:md5",
     "parentField": "cr:parentField",
     "path": "cr:path",
-    "personalSensitiveInformation": "rai:personalSensitiveInformation",
     "recordSet": "cr:recordSet",
     "references": "cr:references",
     "regex": "cr:regex",
@@ -44,17 +41,7 @@
     "source": "cr:source",
     "subField": "cr:subField",
     "transform": "cr:transform",
-    "rai": "http://mlcommons.org/croissant/RAI/",
-    "dataLimitations": "rai:dataLimitations",
-    "dataUseCases": "rai:dataUseCases",
-    "dataSocialImpact": "rai:dataSocialImpact",
-    "hasSyntheticData": "rai:hasSyntheticData",
-    "dataPreprocessing": "rai:dataPreprocessing",
-    "dataAnnotationProtocol": "rai:dataAnnotationProtocol",
-    "safetyMeasures": "rai:safetyMeasures",
-    "deidentificationMethod": "rai:deidentificationMethod",
-    "fairness": "rai:fairness",
-    "releaseMaintenance": "rai:releaseMaintenance"
+    "rai": "http://mlcommons.org/croissant/RAI/"
   },
   "@type": "sc:Dataset",
   "distribution": [
@@ -11126,20 +11113,20 @@
   ],
   "license": "https://choosealicense.com/licenses/apache-2.0/",
   "url": "https://huggingface.co/datasets/PINNBench/pinnbench",
-  "dataCollection": "All result records are derived from controlled training runs of hybrid PINN-operator pipelines on synthetic PDE problems with analytical reference solutions. No human subjects, no scraped web content, no third-party datasets are involved. Runs were executed deterministically with fixed random seeds on 4x NVIDIA TITAN Xp GPUs (12 GB each). Each run records final-epoch evaluation metrics (relative L2 error, PDE-residual norm, BC/IC violations) along with full training history at probe epochs.",
-  "dataPreprocessing": "Reference solutions are computed from analytical PDE expressions (Fourier series for Heat1D, Cole-Hopf for Burgers1D, soliton sech^2 for KdV1D, tanh-front for AllenCahn1D, Taylor-Green vortex for NavierStokes2D, etc.). Training collocation points and evaluation points are sampled per seed via torch.rand with deterministic state. No filtering, normalization, or transformation is applied to recorded metrics; raw per-seed values are persisted.",
-  "dataAnnotationProtocol": "No human-provided annotations. Each run is automatically labeled with its (PDE, configuration, policy, seed, stage) tuple at logging time. Statistical aggregations (mean, std, paired Wilcoxon p-value, Cohen's d, BH-FDR corrected significance) are computed by experiments/scripts/statistical_analysis.py.",
-  "personalSensitiveInformation": "None. The benchmark consists of synthetic numerical-PDE simulations with analytical reference solutions. No personal, identifying, biometric, behavioural, location, health, or other sensitive data is collected, processed, or distributed. No human subjects, no scraped web data, no third-party datasets are involved. The dataset poses zero individual privacy risk by construction and contains no information that could be used to identify any natural person.",
-  "safetyMeasures": "PDE solvers are abstract numerical methodology. Outputs are scalar error metrics on synthetic problems and pose no direct safety risk. Downstream users planning to apply hybrid PINN-operator solvers in safety-critical engineering (climate, fluid simulation, materials) should validate model fidelity and out-of-distribution robustness before deployment; the benchmark documents specific OOD-degradation findings (hybrid models degrade proportionally more than PINN-only on every evaluated coefficient/domain/IC shift) that practitioners should heed.",
-  "deidentificationMethod": "Not applicable. No identifying information present in source data.",
-  "fairness": "The benchmark deliberately reports negative and conditional findings (causal loss helps with pretraining and hurts without it; routing accuracy and regret dissociate by 77x; diagnostic bound fails universally at probe window k=20 and on Burgers-type shock dynamics). Selection of PDEs spans diffusion, advection, nonlinear waves, dispersive, reaction-diffusion, and Navier-Stokes regimes to reduce family-specific bias. Family-level holdout (LOPO-F) is reported.",
-  "releaseMaintenance": "Version 1.0.0 frozen at 2026-05-04 submission time. Single-maintainer issue tracker on the deanonymized GitHub repository (revealed at camera-ready). 12-month issue/PR response SLA. Forward compatibility: pinned PyTorch 2.5.1 + Python 3.12. Planned updates include additional PDE configurations (community PRs invited) and a noisy/sparse-measurement extension.",
-  "dataLimitations": "The benchmark suite contains 13 PDE configurations across 8 equation families and is dominated by 1D problems with a single 2D Wave equation and a single 2D Navier-Stokes configuration. Stage-1 supervised reference data is analytical and noise-free; the benchmark does not measure noisy-measurement, sparse-observation, or real-instrument-data regimes. All training is performed for a fixed compute budget on a single 4xNVIDIA TITAN Xp node; results may not generalise to substantially larger architectures or different optimiser families. Routing meta-features include hand-coded PDE structural attributes that may partly proxy PDE family identity in this small suite, although the family-macro physics-loss-final result (81.2 percent, no PDE identity features used) partially mitigates this concern.",
-  "dataBiases": "By construction the benchmark draws PDEs from diffusion, advection, nonlinear waves, dispersive, reaction-diffusion, and Navier-Stokes regimes to reduce family-specific bias, but: (a) it over-represents 1D problems; (b) reference solutions are analytical exact-solutions, which biases methods that exploit smooth ground truth; (c) the routing benchmark records baseline-policy decisions for a fixed library of five training policies, which biases evaluation toward routing methods that generalise within this policy set. We deliberately publish negative and conditional findings to expose these biases rather than hide them.",
-  "dataUseCases": "Intended use cases: (i) benchmarking training-policy selection methods (probe-then-commit routers, meta-routers, contextual bandits, learning-curve extrapolators) in hybrid PINN-operator pipelines; (ii) controlled PDE-level ablations of pretraining, causal loss, physics-weight balance, and gating-architecture choices; (iii) studying the accuracy-regret dissociation in evaluation-by-accuracy versus evaluation-by-regret. Out-of-scope: production PDE solving, safety-critical engineering deployment (climate forecasting, structural engineering, fluid simulation for aerospace), noisy-measurement or sparse-observation problems, and any clinical, financial, or surveillance application.",
-  "dataSocialImpact": "The dataset itself has minimal direct societal impact: it is purely synthetic numerical-PDE simulation data with no link to individuals, demographics, or real-world systems. Indirect impact arises through the methods that the benchmark advances. PDE solvers based on physics-informed and operator-learning methods feed into safety-critical applications such as climate forecasting, fluid simulation in aerospace and energy, and structural materials engineering, so routing recommendations or training-policy advice derived from this benchmark must be validated out-of-distribution before any real-world deployment. The benchmark explicitly documents OOD-degradation findings (hybrid models degrade proportionally more than PINN-only on every evaluated coefficient/domain/IC shift) to discourage uncritical transfer.",
-  "hasSyntheticData": true,
   "version": "1.0.0",
   "datePublished": "2026-05-04",
-  "citeAs": "Anonymous. PINNBench: A Benchmark and Evaluation Study of Training Policy Selection in Hybrid PINN-Operator Solvers. Submitted to NeurIPS 2026 Evaluations and Datasets Track, 2026."
+  "citeAs": "Anonymous. PINNBench: A Benchmark and Evaluation Study of Training Policy Selection in Hybrid PINN-Operator Solvers. Submitted to NeurIPS 2026 Evaluations and Datasets Track, 2026.",
+  "rai:dataLimitations": "The benchmark suite contains 13 PDE configurations across 8 equation families and is dominated by 1D problems with a single 2D Wave equation and a single 2D Navier-Stokes configuration. Stage-1 supervised reference data is analytical and noise-free; the benchmark does not measure noisy-measurement, sparse-observation, or real-instrument-data regimes. All training is performed for a fixed compute budget on a single 4xNVIDIA TITAN Xp node; results may not generalise to substantially larger architectures or different optimiser families. Routing meta-features include hand-coded PDE structural attributes that may partly proxy PDE family identity in this small suite, although the family-macro physics-loss-final result (81.2 percent, no PDE identity features used) partially mitigates this concern.",
+  "rai:dataBiases": "By construction the benchmark draws PDEs from diffusion, advection, nonlinear waves, dispersive, reaction-diffusion, and Navier-Stokes regimes to reduce family-specific bias, but: (a) it over-represents 1D problems; (b) reference solutions are analytical exact-solutions, which biases methods that exploit smooth ground truth; (c) the routing benchmark records baseline-policy decisions for a fixed library of five training policies, which biases evaluation toward routing methods that generalise within this policy set. We deliberately publish negative and conditional findings to expose these biases rather than hide them.",
+  "rai:personalSensitiveInformation": "None. The benchmark consists of synthetic numerical-PDE simulations with analytical reference solutions. No personal, identifying, biometric, behavioural, location, health, or other sensitive data is collected, processed, or distributed. No human subjects, no scraped web data, no third-party datasets are involved. The dataset poses zero individual privacy risk by construction and contains no information that could be used to identify any natural person.",
+  "rai:dataUseCases": "Intended use cases: (i) benchmarking training-policy selection methods (probe-then-commit routers, meta-routers, contextual bandits, learning-curve extrapolators) in hybrid PINN-operator pipelines; (ii) controlled PDE-level ablations of pretraining, causal loss, physics-weight balance, and gating-architecture choices; (iii) studying the accuracy-regret dissociation in evaluation-by-accuracy versus evaluation-by-regret. Out-of-scope: production PDE solving, safety-critical engineering deployment (climate forecasting, structural engineering, fluid simulation for aerospace), noisy-measurement or sparse-observation problems, and any clinical, financial, or surveillance application.",
+  "rai:dataSocialImpact": "The dataset itself has minimal direct societal impact: it is purely synthetic numerical-PDE simulation data with no link to individuals, demographics, or real-world systems. Indirect impact arises through the methods that the benchmark advances. PDE solvers based on physics-informed and operator-learning methods feed into safety-critical applications such as climate forecasting, fluid simulation in aerospace and energy, and structural materials engineering, so routing recommendations or training-policy advice derived from this benchmark must be validated out-of-distribution before any real-world deployment. The benchmark explicitly documents OOD-degradation findings (hybrid models degrade proportionally more than PINN-only on every evaluated coefficient/domain/IC shift) to discourage uncritical transfer.",
+  "rai:hasSyntheticData": true,
+  "rai:dataCollection": "All result records are derived from controlled training runs of hybrid PINN-operator pipelines on synthetic PDE problems with analytical reference solutions. No human subjects, no scraped web content, no third-party datasets are involved. Runs were executed deterministically with fixed random seeds on 4x NVIDIA TITAN Xp GPUs (12 GB each). Each run records final-epoch evaluation metrics (relative L2 error, PDE-residual norm, BC/IC violations) along with full training history at probe epochs.",
+  "rai:dataPreprocessing": "Reference solutions are computed from analytical PDE expressions (Fourier series for Heat1D, Cole-Hopf for Burgers1D, soliton sech^2 for KdV1D, tanh-front for AllenCahn1D, Taylor-Green vortex for NavierStokes2D, etc.). Training collocation points and evaluation points are sampled per seed via torch.rand with deterministic state. No filtering, normalization, or transformation is applied to recorded metrics; raw per-seed values are persisted.",
+  "rai:dataAnnotationProtocol": "No human-provided annotations. Each run is automatically labeled with its (PDE, configuration, policy, seed, stage) tuple at logging time. Statistical aggregations (mean, std, paired Wilcoxon p-value, Cohen's d, BH-FDR corrected significance) are computed by experiments/scripts/statistical_analysis.py.",
+  "rai:safetyMeasures": "PDE solvers are abstract numerical methodology. Outputs are scalar error metrics on synthetic problems and pose no direct safety risk. Downstream users planning to apply hybrid PINN-operator solvers in safety-critical engineering (climate, fluid simulation, materials) should validate model fidelity and out-of-distribution robustness before deployment; the benchmark documents specific OOD-degradation findings (hybrid models degrade proportionally more than PINN-only on every evaluated coefficient/domain/IC shift) that practitioners should heed.",
+  "rai:deidentificationMethod": "Not applicable. No identifying information present in source data.",
+  "rai:fairness": "The benchmark deliberately reports negative and conditional findings (causal loss helps with pretraining and hurts without it; routing accuracy and regret dissociate by 77x; diagnostic bound fails universally at probe window k=20 and on Burgers-type shock dynamics). Selection of PDEs spans diffusion, advection, nonlinear waves, dispersive, reaction-diffusion, and Navier-Stokes regimes to reduce family-specific bias. Family-level holdout (LOPO-F) is reported.",
+  "rai:releaseMaintenance": "Version 1.0.0 frozen at 2026-05-04 submission time. Single-maintainer issue tracker on the deanonymized GitHub repository (revealed at camera-ready). 12-month issue/PR response SLA. Forward compatibility: pinned PyTorch 2.5.1 + Python 3.12. Planned updates include additional PDE configurations (community PRs invited) and a noisy/sparse-measurement extension."
 }