README.md

---
license: cc-by-4.0
tags:
  - calibration
  - post-hoc calibration
  - uncertainty quantification
  - benchmark
  - tabular
  - computer vision
size_categories:
  - 1M<n<10M
---

# CalArena — Calibration Benchmark Dataset

CalArena is a large-scale benchmark for evaluating post-hoc calibration methods on classification models.
It covers **7 benchmarks** across tabular and computer vision domains, spanning hundreds of (dataset, model) pairs and three problem types (binary, multiclass and large scale multiclass).

Each entry in the benchmark is a `(p_cal, y_cal, p_test, y_test)` tuple — the calibration split and test split of predicted probabilities and ground-truth labels for one (dataset, model) pair.
Calibration methods are fitted on the calibration split and evaluated on the test split.

This dataset is the data companion to the [CalArena code repository](https://github.com/super-anonymous-researcher/CalArena).

---

## Files

| File | Description | Size |
|---|---|---|
| `Licenses.zip` | License files for each data source used to create the benchmarks | < 1 MB |
| `tabrepo-binary.h5` | Binary classification, classical tabular models | ~36 MB |
| `tabrepo-binary-experiments.csv` | Experiment index for `tabrepo-binary` | < 1 MB |
| `tabarena-binary.h5` | Binary classification, modern tabular foundation models | ~26 MB |
| `tabarena-binary-experiments.csv` | Experiment index for `tabarena-binary` | < 1 MB |
| `cv-binary.h5` | Binary classification, computer vision models | < 1 MB |
| `cv-binary-experiments.csv` | Experiment index for `cv-binary` | < 1 MB |
| `tabrepo-multiclass.h5` | Multiclass classification, classical tabular models | ~115 MB |
| `tabrepo-multiclass-experiments.csv` | Experiment index for `tabrepo-multiclass` | < 1 MB |
| `tabarena-multiclass.h5` | Multiclass classification, modern tabular foundation models | ~11 MB |
| `tabarena-multiclass-experiments.csv` | Experiment index for `tabarena-multiclass` | < 1 MB |
| `cv-multiclass.h5` | Multiclass classification, computer vision models | ~39 MB |
| `cv-multiclass-experiments.csv` | Experiment index for `cv-multiclass` | < 1 MB |
| `imagenet-multiclass.h5` | 1000-class ImageNet, computer vision models | ~1.5 GB |
| `imagenet-multiclass-experiments.csv` | Experiment index for `imagenet-multiclass` | < 1 MB |

---

## Benchmark overview

| Benchmark | Problem type | Base models | # Datasets | # Experiments |
|---|---|---|---|---|
| `tabrepo-binary` | Binary | 8 | 104 tabular datasets | 832 |
| `tabarena-binary` | Binary | 11 | 30 tabular datasets | 314 |
| `cv-binary` | Binary | 9 | 3 (CIFAR-10†, Breast, Pneumonia) | 13 |
| `tabrepo-multiclass` | Multiclass | 8 | 65 tabular datasets | 520 |
| `tabarena-multiclass` | Multiclass | 11 | 8 tabular datasets | 84 |
| `cv-multiclass` | Multiclass | 10 | 6 (CIFAR-10, CIFAR-100, Birds, SVHN, Derma, OCT) | 20 |
| `imagenet-multiclass` | Large scale multiclass | 8 | 1 (ImageNet) | 8 |

† CIFAR-10 is converted to binary (Animal vs Machine) by marginalising over class groups.

### Base models

**TabRepo** (classical tabular): CatBoost, ExtraTrees, LightGBM, LinearModel, NeuralNetFastAI, NeuralNetTorch, RandomForest, XGBoost. Source: [TabRepo](https://github.com/autogluon/tabrepo) repository `D244_F3_C1530_200`.
Best hyperparameter configuration selected per (dataset, model, fold) by validation error.

**TabArena** (modern tabular): TabPFN-v2.6, TabICLv2, RealTabPFN-v2.5, TabICL\_GPU, LimiX\_GPU, TabM\_GPU, RealMLP\_GPU, BetaTabPFN\_GPU, ModernNCA\_GPU, Mitra\_GPU, TabDPT\_GPU.
Models selected with ≥ 1300 ELO on the [TabArena leaderboard](https://huggingface.co/spaces/TabArena/leaderboard) (Classification, All Datasets, as of April 1 2026).
Source: [TabArena](https://github.com/autogluon/tabarena).

**Computer vision**: ResNet, DenseNet, WideResNet, ViT, BEiT, ConvNeXt, Swin, EVA, and others depending on the dataset.
Logits sourced from two collections: [NN_calibration](https://github.com/markus93/NN_calibration/tree/master/logits) and [Beyond Overconfidence](https://zenodo.org/records/15229730).

---

## Data format

### HDF5 files

Each `.h5` file has the following structure:

```
{dataset}/
  {model}/
    probas_cal   float32  (n_cal,)           # positive-class probabilities [binary]
                 float32  (n_cal, n_classes) # class probabilities [multiclass]
    labels_cal   int32    (n_cal,)
    probas_test  float32  (n_test,)          # same shape conventions as above
    labels_test  int32    (n_test,)
```

File-level attributes:
- `source` — `"tabrepo"`, `"tabarena"`, `"cv"`, or `"imagenet"`
- `problem_type` — `"binary"` or `"multiclass"`

All probabilities are valid (non-negative, sum to 1 for multiclass). Labels are 0-indexed integers.

### Experiment CSV files

Each `{benchmark}-experiments.csv` lists one row per (dataset, model) pair:

| Column | Description |
|---|---|
| `dataset` | Dataset name (matches the HDF5 group key) |
| `model` | Model name (matches the HDF5 group key) |
| `cal_size` | Number of calibration samples |
| `test_size` | Number of test samples |
| `n_classes` | Number of classes (multiclass benchmarks only) |
| `tabrepo_fold` / `tabarena_fold` | Fold index used (TabRepo/TabArena benchmarks) |
| `tabrepo_config` / `tabarena_config` | Best hyperparameter configuration selected (TabRepo/TabArena) |

---

## Loading the data

### Python (h5py)

```python
import h5py
import numpy as np

with h5py.File("tabrepo-binary.h5", "r") as f:
    # List all (dataset, model) pairs
    pairs = [(ds, mdl) for ds in f for mdl in f[ds]]

    # Load a single experiment
    grp = f["anneal/CatBoost"]
    p_cal   = grp["probas_cal"][:]   # shape (n_cal,)
    y_cal   = grp["labels_cal"][:]   # shape (n_cal,)
    p_test  = grp["probas_test"][:]  # shape (n_test,)
    y_test  = grp["labels_test"][:]  # shape (n_test,)
```

### With the CalArena runner

The [CalArena repository](https://github.com/super-anonymous-researcher/CalArena) provides `run_benchmark.py`, which loads these files automatically and runs all calibrators:

```bash
# Place .h5 and .csv files under calibration_benchmarks/
python run_benchmark.py --benchmark tabrepo-binary
```

---

## Dataset construction

Scripts that were used to generate the benchmarks files can be found in the [CalArena repository](https://github.com/super-anonymous-researcher/CalArena).

### Calibration / test split

For TabRepo and TabArena, the calibration split corresponds to the **validation fold** of the respective repository, and the test split is the **held-out test set**.
This ensures no data leakage: the base model never sees the calibration set during training.

For computer vision datasets, the calibration and test splits are fixed partitions provided by the original data sources.

### Excluded datasets

The following datasets were excluded due to errors in the upstream repositories:

- **TabRepo binary**: MiniBooNE
- **TabRepo multiclass**: jannis, kropt, shuttle

---

## Intended use

This dataset is intended for:
- Benchmarking post-hoc calibration algorithms on diverse classification tasks
- Studying the relationship between model type, dataset characteristics, and calibration difficulty
- Developing new calibration methods with access to pre-computed probability estimates

---

## License

The benchmark data is released under **CC BY 4.0**.
Downstream sources of model predictions retain their original licenses; please consult the respective sources before redistribution:
- [TabArena](https://github.com/autogluon/tabarena)
- [TabRepo](https://github.com/autogluon/tabarena/blob/main/tabrepo.md)
- [NN_calibration](https://github.com/markus93/NN_calibration)
- [Data for: "Beyond Overconfidence..."](https://zenodo.org/records/15229730)

We warmly thank the authors of the original papers for letting us republish their model predictions here.

---

## Citation

```bibtex
@inproceedings{calarena2025,
  title     = {CalArena: A Large-Scale Benchmark for Post-Hoc Calibration},
  author    = {...},
  booktitle = {...},
  year      = {2025},
}
```