| --- |
| license: mit |
| metrics: |
| - accuracy |
| - roc_auc |
| - demographic_parity |
| - equalized_odds |
| - equal_opportunity |
| pipeline_tag: tabular-classification |
| tags: |
| - tabular |
| - fairness |
| - TFM |
| base_model: |
| - FairTFM-0.7-epoch_10000.pt |
| --- |
| # Model Card: FairTFM |
|
|
| ## Model Description |
| FairTFM is a **tabular foundation model** for fairness-aware, in-context prediction on tabular data. It is a transformer-based architecture (built on components from [nanoTabPFN](https://github.com/automl/TFM-Playground)) that jointly embeds features, targets, and a **sensitive attribute** via a dedicated `SensitiveAttributeEncoder`, then predicts labels for unseen (test) rows conditioned on a small set of in-context training rows — no per-dataset fine-tuning required. |
|
|
| - **Architecture:** Transformer encoder stack with separate feature/target/sensitive-attribute embeddings, decoded to per-row logits. Memory-chunked attention/MLP for efficient inference. |
| - **Task type:** Binary classification with a fairness objective over a specified sensitive attribute. |
| - **Release type:** Inference-only. |
|
|
| ### Model Sources |
| |
|
|
| - **Repository:** https://github.com/patrikken/FairTFM-inference |
| - **Paper:** https://openreview.net/forum?id=ajIvCEbadL |
|
|
| ## Checkpoints |
| Four checkpoints are provided, corresponding to different fairness-regularization strengths (λ) used to trace a fairness/accuracy Pareto front: **λ = 0.7, 1.0, 10, 25**. Higher λ trades predictive performance for lower fairness-metric disparity. |
|
|
| ## Intended Use |
| - Fair classification on tabular datasets where a sensitive attribute (e.g., sex, race, age) is known at both train and inference time. |
| - Includes benchmark support for the folktables **ACS PUMS** tasks (income, employment, mobility, public coverage, travel time), among other tabular fairness datasets. |
| - Sensitive attributes evaluated: `SEX`, `RAC1P` (race, White/Black), `AGEP` (age, median-binarized). |
|
|
|
|
| ## Limitations |
| - Restricted to binary sensitive attributes and binary classification targets in current released checkpoints. |
| - Fairness metrics are computed with respect to the single sensitive attribute provided at inference; intersectional fairness is not directly modeled. |
|
|
| ## How to Use |
|
|
| ```python |
| from fairtfm import FairTFMClassifier, compute_fairness_metrics |
| |
| # Load checkpoint |
| classifier = FairTFMClassifier(model="path/to/checkpoint") |
| |
| # Fit on training data |
| classifier.fit(X_train, y_train, s_train) |
| |
| # Predict |
| predictions = classifier.predict(X_test, s_test) |
| probabilities = classifier.predict_proba(X_test, s_test) |
| |
| # Fairness metrics (returns dict with performance metrics) |
| compute_fairness_metrics(X_test, y_test, s_test) |
| ``` |
|
|
| ## Citation |
| ```bibtex |
| @inproceedings{ |
| kenfack2026training, |
| title={Training Fair Tabular Foundation Models}, |
| author={Patrik Kenfack and Jesse C. Cresswell and Anthony L. Caterini and Samira Ebrahimi Kahou and Ulrich A{\"\i}vodji}, |
| booktitle={2nd ICML Workshop on Foundation Models for Structured Data}, |
| year={2026}, |
| url={https://openreview.net/forum?id=ajIvCEbadL} |
| } |
| ``` |
