Datasets:

EvalData
/

TabMI-Bench

Purpose: TabMI-Bench provides a standardized protocol for evaluating mechanistic interpretability (MI) methods on Tabular Foundation Models (TFMs). It fills the gap between mature LLM MI evaluation (e.g., MIB) and the growing but uninterpreted TFM ecosystem.
Creators: Anonymous (double-blind submission). Will be de-anonymized upon acceptance.
Funding: [To be disclosed upon acceptance]
Date of creation: January–April 2026

2. Composition

Benchmark Components

Component	Description	Files	Format
Hook infrastructure	Activation extraction for 5 TFMs	`src/hooks/*.py` (10 files)	Python
Synthetic data generators	4 function types (bilinear, sinusoidal, polynomial, mixed)	`src/data/synthetic_generator.py`	Python
Real-world dataset loaders	11 causal tracing + 4 steering datasets	`src/data/real_world_datasets.py`	Python (scikit-learn, OpenML)
Evaluation scripts	Experiment scripts for all MI techniques	`experiments/*.py`	Python
Reference baselines	Multi-seed results for 5 models	`results/`	JSON
Negative controls	Random-target, shuffled-label, random-vector	`experiments/neurips_c1c2_baselines_*.py`	Python
Benchmark card	This document	`BENCHMARK_CARD.md`	Markdown
Croissant metadata	Machine-readable metadata	`croissant.json`	JSON-LD
Example: add new model	Worked extensibility example	`examples/add_new_model.py`	Python

Instance Counts

Synthetic probes: 4 function families × configurable N (default N_train=100, N_test=50)
Real-world datasets: 11 for causal tracing, 4 for steering (all public, from scikit-learn/OpenML)
Models: 5 TFMs (TabPFN v2, TabPFN v2.5, TabICL v2, TabDPT, iLTM)
MI techniques: 8 (linear probing, logit lens, activation patching, vector ablation, steering vectors, SAE, CKA, attention analysis)
Reference seeds: 5 (core probing on TabPFN/TabICL/iLTM), 3 (TabDPT in-family holdout, N=10K probing scale, real-world causal), 5 (NAM out-of-family holdout, real-world steering, function invariance), 10 (robust steering, SAE expansion)

Data Types

Synthetic: continuous numeric features only (no categorical, no missing values)
Real-world: mixed (numeric + categorical with encoding; no explicit missing value handling in synthetic core)

Sensitive Information

No personally identifiable information (PII) in synthetic data
Real-world datasets inherit properties of their public sources:
- Adult Income (OpenML 1590): contains age, sex, race attributes
- California Housing: geographic data reflecting historical patterns
The benchmark evaluates MI methods, not model fairness

Instances Not Included

Concrete dataset excluded due to OpenML cache failure
Closed-source TFMs excluded (no checkpoint access)
Non-ICL tabular methods (XGBoost, LightGBM) excluded by design

3. Collection Process

Synthetic data: Generated deterministically from mathematical functions with fixed random seeds. No human data collection.
Real-world data: Loaded from scikit-learn built-in datasets and OpenML public repositories at runtime. No new data collection.
Model activations: Extracted via forward hooks registered on PyTorch modules during inference. No model training performed.
Timeframe: Experiments conducted January–March 2026

4. Preprocessing

Synthetic: Features generated from known distributions; targets computed from closed-form functions with Gaussian noise.
Real-world: StandardScaler applied to features (zero mean, unit variance). Train/test split with fixed random seed.
Activations: Raw layer outputs stored as float64 numpy arrays. No dimensionality reduction applied before probing.

5. Uses

Intended Uses

Evaluating new MI techniques: Run the 4-step protocol, compare results against reference baselines
Profiling new TFMs: Implement hook interface (~150 lines), run existing scripts
Educational: Understanding how different TFM architectures process tabular data internally
Reproducibility: Independent verification of the reported computation profiles

Not Intended For

Deployment certification: Benchmark outputs are diagnostic measurements, NOT safety/fairness/compliance certifications
Model ranking: TabMI-Bench is a protocol benchmark, not a scored leaderboard
Prediction accuracy comparison: Use TabZilla, TabReD, or OpenML-CC18 for predictive performance
Security auditing: The benchmark does not test adversarial robustness or data poisoning

Potential Misuse

Over-interpreting "staged" or "distributed" labels as complete model understanding
Using steering controllability results as evidence of safe/aligned controllability
Treating benchmark pass as regulatory compliance evidence

6. Distribution

License: MIT (see LICENSE file)
Hosting:
- Code and scripts: GitHub (anonymous during review, public upon acceptance)
- Archived artifacts: Zenodo (DOI-minted for long-term persistence)
- ML-ecosystem discovery: HuggingFace Datasets (Croissant-annotated)
Access: Open access. No registration required for code or synthetic data. Real-world datasets require internet for OpenML downloads.
Export controls: None. No dual-use concerns.

7. Maintenance

Maintainers: Paper authors (contact: [upon acceptance])
Update policy: Version-pinned snapshots for reproducibility. New model hooks and reference baselines added as TFMs are released.
Versioning: Semantic versioning (major.minor.patch). Breaking changes increment major version.
Community contributions: Pull requests welcomed for new model hook implementations
Deprecation: Individual model hooks may be deprecated when models are superseded, but archived results are preserved

8. Known Limitations

Limitation	Scope	Mitigation
Small scale (N=100)	Core experiments	N=10K scale check preserves patterns
Regression only (synthetic)	Core probes	Classification probing in appendix
5 models, 3 families	Taxonomy scope	Holdout validation on TabDPT (3-seed) and NAM (5-seed); extensible hook design
Post-hoc taxonomy	Scientific claim	Profile variance metric + ±50% threshold robustness + leave-one-function-out (LOFO) verified primary-endpoint ordering with ratio >29× in every holdout
N=10K steering sub-check is single-seed	Evidence tier	Probing component uses 3 seeds; steering sub-check labeled supporting
No categorical synthetic features	Probe design	Real-world datasets include categorical
Probing collapses at d≥8	Method limitation	Causal tracing remains informative

9. Ethical Considerations

The benchmark studies existing open-source models and does not create new capabilities
Real-world datasets are publicly available and widely used in ML research
Steering demonstrations should NOT be interpreted as evidence of safe model controllability
Users should not rely on benchmark outputs for regulatory compliance without additional validation

10. Citation

@inproceedings{anonymous2026tabmibench,
  title={TabMI-Bench: Evaluating Mechanistic Interpretability Methods Across Tabular Foundation Model Architectures},
  author={Anonymous},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}