{ "@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/", "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" }, "@type": "sc:Dataset", "name": "TabMI-Bench", "description": "A protocol benchmark for evaluating mechanistic interpretability (MI) methods on Tabular Foundation Models (TFMs). Provides hook-based activation extraction for 5 TFMs across 3 architectural families, controlled synthetic probes (4 function types), negative controls, a 4-step evaluation protocol, and reference baseline measurements.", "conformsTo": "http://mlcommons.org/croissant/1.0", "url": "https://anonymous.4open.science/r/TabMI-Bench/", "sameAs": "https://huggingface.co/datasets/EvalData/TabMI-Bench", "license": "https://opensource.org/licenses/MIT", "version": "1.0.0", "datePublished": "2026-04-15", "creator": { "@type": "sc:Organization", "name": "Anonymous (double-blind submission)" }, "keywords": [ "mechanistic interpretability", "tabular foundation models", "benchmark", "probing", "activation patching", "sparse autoencoders", "steering vectors" ], "inLanguage": "en", "citeAs": "@inproceedings{anonymous2026tabmibench, title={TabMI-Bench: Evaluating Mechanistic Interpretability Methods Across Tabular Foundation Model Architectures}, author={Anonymous}, booktitle={Advances in Neural Information Processing Systems (NeurIPS) Evaluations \\& Datasets Track}, year={2026}}", "distribution": [ { "@type": "cr:FileObject", "@id": "rd5-fullscale-aggregated", "name": "rd5-fullscale-aggregated", "description": "Aggregated 5-seed multi-model intermediary probing results (Table 1 source).", "contentUrl": "https://huggingface.co/datasets/EvalData/TabMI-Bench/resolve/main/rd5_fullscale_aggregated.json", "encodingFormat": "application/json", "sha256": "8bae47fa6f777ff1b8e4d8950470d99f3a024478d43997b09f4b24642881a1c3" }, { "@type": "cr:FileObject", "@id": "tabdpt-probing-3seed", "name": "tabdpt-probing-3seed", "description": "TabDPT in-family holdout probing results aggregated over 3 seeds.", "contentUrl": "https://huggingface.co/datasets/EvalData/TabMI-Bench/resolve/main/tabdpt_probing_3seed.json", "encodingFormat": "application/json", "sha256": "09e6fe1cfcdec668ee6bbc0f8927aef63af4f10039c7db28145b4646ee53f055" }, { "@type": "cr:FileObject", "@id": "tabdpt-causal-3seed", "name": "tabdpt-causal-3seed", "description": "TabDPT noising-based causal tracing results aggregated over 3 seeds.", "contentUrl": "https://huggingface.co/datasets/EvalData/TabMI-Bench/resolve/main/tabdpt_causal_3seed.json", "encodingFormat": "application/json", "sha256": "a19828d2f5ca3c3812651d5179532ff3fa882a01d92cf74735cab195859cf54b" }, { "@type": "cr:FileObject", "@id": "nam-holdout", "name": "nam-holdout", "description": "NAM out-of-family holdout per-seed layer-wise R^2 profiles.", "contentUrl": "https://huggingface.co/datasets/EvalData/TabMI-Bench/resolve/main/nam_holdout.json", "encodingFormat": "application/json", "sha256": "ebf795a470ea11e5fa5b6d2a85c6287add70494cecba81e94ba038ed2c3653af" }, { "@type": "cr:FileObject", "@id": "lofo-primary-endpoint", "name": "lofo-primary-endpoint", "description": "Leave-one-function-out robustness for the primary endpoint sigma2_profile.", "contentUrl": "https://huggingface.co/datasets/EvalData/TabMI-Bench/resolve/main/lofo_primary_endpoint.json", "encodingFormat": "application/json", "sha256": "66fe13294502cd9c0a32616861b0ccb4aa11e7544da54f60b98943ad55173419" }, { "@type": "cr:FileObject", "@id": "tabpfn25-fullscale-aggregated", "name": "tabpfn25-fullscale-aggregated", "description": "TabPFN v2 vs v2.5 logit-lens, coefficient probing, intermediary probing, and CKA aggregated over 3 seeds.", "contentUrl": "https://huggingface.co/datasets/EvalData/TabMI-Bench/resolve/main/tabpfn25_fullscale_aggregated.json", "encodingFormat": "application/json", "sha256": "3f81492cd2183a5e8bc82216162c75579bffccf5c19a0cc07b7533a85db94773" }, { "@type": "cr:FileObject", "@id": "c1c2-baselines", "name": "c1c2-baselines", "description": "Negative-control baselines: shuffled-label probing and random-target probing per model and layer.", "contentUrl": "https://huggingface.co/datasets/EvalData/TabMI-Bench/resolve/main/c1c2_baselines.json", "encodingFormat": "application/json", "sha256": "c19cd33ef8e1e7b03ad26956cff8f932d985ce4f1d234a030d619bd7a7aadf5f" }, { "@type": "cr:FileObject", "@id": "scale-10k-multiseed", "name": "scale-10k-multiseed", "description": "N=10K probing scale validation results aggregated over 3 seeds.", "contentUrl": "https://huggingface.co/datasets/EvalData/TabMI-Bench/resolve/main/scale_10k_multiseed.json", "encodingFormat": "application/json", "sha256": "ecf8cb8572dcd9f4ea1c6cd569b52ab9b38a2d406ab18bb1b2da766f8e853eca" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "synthetic-probing-results", "name": "synthetic-probing-results", "description": "Layer-wise intermediary probing R^2 values for each (model, seed, function_type, layer). Source: rd5_fullscale_aggregated.json (intermediary_probing block).", "field": [ { "@type": "cr:Field", "@id": "synthetic-probing-results/model", "name": "model", "description": "TFM model name (one of: tabpfn, tabicl, tabdpt, iltm, nam).", "dataType": "sc:Text", "source": { "fileObject": {"@id": "rd5-fullscale-aggregated"}, "extract": {"jsonPath": "$.intermediary_probing.*"} } }, { "@type": "cr:Field", "@id": "synthetic-probing-results/seed", "name": "seed", "description": "Random seed (one of: 42, 123, 456, 789, 1024).", "dataType": "sc:Integer", "source": { "fileObject": {"@id": "rd5-fullscale-aggregated"}, "extract": {"jsonPath": "$.seeds[*]"} } }, { "@type": "cr:Field", "@id": "synthetic-probing-results/function_type", "name": "function_type", "description": "Synthetic function family (bilinear, sinusoidal, polynomial, mixed).", "dataType": "sc:Text", "source": { "fileObject": {"@id": "rd5-fullscale-aggregated"}, "extract": {"jsonPath": "$.function_type"} } }, { "@type": "cr:Field", "@id": "synthetic-probing-results/layer", "name": "layer", "description": "Layer index (0-based).", "dataType": "sc:Integer", "source": { "fileObject": {"@id": "rd5-fullscale-aggregated"}, "extract": {"jsonPath": "$.intermediary_probing.tabpfn.intermediary_r2_mean[*]"} } }, { "@type": "cr:Field", "@id": "synthetic-probing-results/intermediary_r2", "name": "intermediary_r2", "description": "Linear-probe R^2 for the intermediary variable at this (model, seed, function, layer).", "dataType": "sc:Float", "source": { "fileObject": {"@id": "rd5-fullscale-aggregated"}, "extract": {"jsonPath": "$.intermediary_probing.*.intermediary_r2_mean[*]"} } } ] }, { "@type": "cr:RecordSet", "@id": "causal-tracing-results", "name": "causal-tracing-results", "description": "Noising-based causal sensitivity per layer for each (model, seed, layer). Source: tabdpt_causal_3seed.json and rd5_fullscale_aggregated.json patching block.", "field": [ { "@type": "cr:Field", "@id": "causal-tracing-results/model", "name": "model", "description": "TFM model name.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "tabdpt-causal-3seed"}, "extract": {"jsonPath": "$.model"} } }, { "@type": "cr:Field", "@id": "causal-tracing-results/seed", "name": "seed", "description": "Random seed.", "dataType": "sc:Integer", "source": { "fileObject": {"@id": "tabdpt-causal-3seed"}, "extract": {"jsonPath": "$.seeds[*]"} } }, { "@type": "cr:Field", "@id": "causal-tracing-results/layer", "name": "layer", "description": "Layer index (0-based).", "dataType": "sc:Integer", "source": { "fileObject": {"@id": "tabdpt-causal-3seed"}, "extract": {"jsonPath": "$.mean_normalized_sensitivity[*]"} } }, { "@type": "cr:Field", "@id": "causal-tracing-results/normalized_sensitivity", "name": "normalized_sensitivity", "description": "Normalized MSE-increase from layer-wise noise injection (max-normalized within profile).", "dataType": "sc:Float", "source": { "fileObject": {"@id": "tabdpt-causal-3seed"}, "extract": {"jsonPath": "$.mean_normalized_sensitivity[*]"} } } ] }, { "@type": "cr:RecordSet", "@id": "applicability-matrix", "name": "applicability-matrix", "description": "Evidence-coded MI technique applicability labels per model (Table 8 in paper, hand-curated reference table; canonical encoding in BENCHMARK_CARD.md and Table 8 LaTeX source).", "field": [ { "@type": "cr:Field", "@id": "applicability-matrix/technique", "name": "technique", "description": "MI technique name (e.g. linear-probing, logit-lens, activation-patching, sae, cka).", "dataType": "sc:Text", "source": { "fileObject": {"@id": "rd5-fullscale-aggregated"}, "extract": {"jsonPath": "$.technique"} } }, { "@type": "cr:Field", "@id": "applicability-matrix/model", "name": "model", "description": "TFM model name.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "rd5-fullscale-aggregated"}, "extract": {"jsonPath": "$.model"} } }, { "@type": "cr:Field", "@id": "applicability-matrix/label", "name": "label", "description": "Applicability label: Supported, Limited, or Not established.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "rd5-fullscale-aggregated"}, "extract": {"jsonPath": "$.label"} } } ] } ], "rai:dataCollection": "Synthetic data is generated deterministically from known mathematical functions with fixed random seeds. Real-world datasets are sourced from scikit-learn built-in datasets and OpenML public repositories. No human subjects data is collected.", "rai:dataCollectionType": "Synthetic generation + public dataset aggregation", "rai:dataCollectionMissingness": "One real-world dataset (Concrete) was excluded due to an OpenML cache failure. All other datasets loaded successfully.", "rai:dataBiases": "Synthetic probes use continuous numeric features only; categorical features and missing values are not tested in the core benchmark. Real-world datasets inherit biases from their original sources (e.g., California Housing reflects historical housing patterns). The benchmark evaluates MI methods, not model fairness.", "rai:dataLimitations": [ "Core experiments use small scale (N_train=100, N_test=50)", "Only regression synthetic probes; classification probes are secondary", "5 models from 3 architectural families; generalization to future TFMs is not guaranteed", "Computation profiles are descriptive reference baselines, not universal laws", "Benchmark outputs are diagnostic measurements, not deployment certifications" ], "rai:dataUseCases": "Evaluating new MI techniques on TFMs, benchmarking new TFM architectures against MI reference baselines, educational resource for understanding TFM internal computation.", "rai:dataSensitiveInformation": "No personally identifiable information (PII). No sensitive attributes beyond those in the original public datasets (e.g., Adult Income contains demographic attributes).", "rai:personalSensitiveInformation": "No personally identifiable information (PII) is collected, generated, or distributed. Synthetic probes contain no human-subjects data. Real-world datasets are loaded at runtime from public repositories (scikit-learn, OpenML) and inherit only the demographic attributes already public in those sources (e.g., Adult Income age/sex/race columns); these are NOT augmented or transformed by TabMI-Bench. The benchmark itself does not perform any inference on individuals; it analyses model internal activations on tabular inputs.", "rai:dataSocialImpact": "Positive: enables standardized MI evaluation for tabular AI in high-stakes domains (credit scoring, medical diagnosis). Risk: benchmark outputs may create false confidence if over-interpreted as safety certifications. Mitigation: explicit scope limitations and risk caveats in the paper.", "rai:dataSyntheticGeneration": "Core synthetic probes are fully synthetic with known ground-truth intermediary variables. Real-world datasets are not synthetic.", "rai:hasSyntheticData": true, "rai:dataMaintenancePlan": "Version-pinned snapshots ensure reproducibility. New model hooks and reference baselines will be added as TFMs are released. Community contributions welcomed via pull requests.", "rai:dataSourceDatasets": [ {"name": "California Housing", "url": "https://www.openml.org/d/8092", "license": "CC0", "purpose": "real-world causal tracing + steering validation"}, {"name": "Diabetes (sklearn)", "url": "https://scikit-learn.org/stable/datasets/toy_dataset.html", "license": "BSD-3-Clause", "purpose": "real-world causal tracing + steering validation"}, {"name": "Wine Quality", "url": "https://www.openml.org/d/287", "license": "CC0", "purpose": "real-world steering"}, {"name": "Bike Sharing", "url": "https://www.openml.org/d/44063", "license": "CC0", "purpose": "real-world steering"}, {"name": "Abalone", "url": "https://www.openml.org/d/183", "license": "CC0", "purpose": "real-world causal tracing"}, {"name": "Boston Housing", "url": "https://www.openml.org/d/531", "license": "CC0", "purpose": "real-world causal tracing"}, {"name": "Energy Efficiency", "url": "https://www.openml.org/d/42178", "license": "CC0", "purpose": "real-world causal tracing"}, {"name": "Breast Cancer (sklearn)", "url": "https://scikit-learn.org/stable/datasets/toy_dataset.html", "license": "BSD-3-Clause", "purpose": "real-world causal tracing + classification probing"}, {"name": "Iris (sklearn)", "url": "https://scikit-learn.org/stable/datasets/toy_dataset.html", "license": "BSD-3-Clause", "purpose": "classification probing"}, {"name": "Adult Income", "url": "https://www.openml.org/d/1590", "license": "CC0", "purpose": "real-world causal tracing"}, {"name": "Credit-G", "url": "https://www.openml.org/d/31", "license": "CC0", "purpose": "real-world causal tracing"} ], "rai:dataProvenance": { "collection": "All real-world datasets are loaded at runtime from public repositories (scikit-learn built-in datasets, OpenML.org); no new human-subjects data collection performed.", "preprocessing": "Real-world: StandardScaler applied with fixed random seed for train/test split; categorical features encoded via OpenML's default encoders. Synthetic: features generated via numpy.random.default_rng(seed) and labels computed from closed-form mathematical functions with optional Gaussian noise.", "annotation": "No human annotation. Synthetic-probe ground-truth intermediary variables (e.g., a*b for bilinear probe) are computed analytically from input features." } }