README.md

---
license: other
license_name: mixed-source-licenses
license_link: https://github.com/injurise/ICMIL/blob/main/THIRD_PARTY.md
pretty_name: ICMIL Benchmarks
task_categories:
  - tabular-classification
tags:
  - multiple-instance-learning
  - benchmark
---

# ICMIL Benchmark Datasets

📄 [Paper (arXiv:2606.06458)](https://arxiv.org/abs/2606.06458) · 💻 [Code](https://github.com/injurise/ICMIL) · 🤗 [Model](https://huggingface.co/bifold-pathomics/ICMIL)

The twelve multiple-instance learning (MIL) benchmarks used to evaluate
[ICMIL](https://huggingface.co/bifold-pathomics/ICMIL) and the baselines. ICMIL is an
in-context learner that labels new bags of instances from a handful of labelled bags
in a single forward pass, without task-specific training. The train/test splits here
are fixed, so numbers from different methods are comparable.

## Files

| File | Benchmark tasks | Source features |
|---|---|---|
| `uci_benchmark.h5` | `uci_musk1`, `uci_musk2`, `uci_letters`, `uci_hepmass` | UCI raw attributes, PCA to 25 dims (16 for `letters`) |
| `mnist_xai_benchmark_100bags.h5` | `mnist_xai_smil`, `mnist_xai_pos_neg`, `mnist_xai_adjacent_pairs`, `mnist_xai_four_bags` | MNIST digits, ResNet-18 embeddings, PCA to 25 |
| `andrews_mil_benchmark.h5` | `andrews_fox`, `andrews_tiger`, `andrews_elephant` | Blobworld region descriptors (230 dims), PCA to 25 |
| `tcga_uni2_luad_vs_lusc.h5` | `tcga_fixed` | TCGA WSI patches, UNI2 features (1536 dims), PCA to 25 |
| `rsna_ich_resnet50_draws_100bags.h5` | `rsna_ich_draws` | RSNA-ICH CT slices, ResNet-50 features (2048 dims), PCA to 25 |

## Getting started

Each file is an HDF5 file with the train/test splits already made, so you can load a
benchmark and start straight away:

```python
import h5py

with h5py.File("uci_benchmark.h5") as f:
    g = f["musk1"]["split_0"]
    X_train, y_train = g["X_train"][:], g["y_train"][:]
    X_test, y_test = g["X_test"][:], g["y_test"][:]

X_train.shape   # (n_bags, bag_size, n_features)
y_train.shape   # (n_bags,) — one label per bag
```

Finding the splits inside a file:

- They are named `split_0`, `split_1`, … where the split is fixed, and `draw_0`,
  `draw_1`, … where it is one of several random draws.
- In the UCI, Andrews and MNIST-XAI files each task has its own group, so a split
  lives at e.g. `musk1/split_0`. The TCGA and RSNA files each hold a single task, so
  the splits sit at the top level: `split_0`, `draw_0`.
- A file's attributes say which feature extractor and split settings were used.

## Licensing

| File | Source | Licence |
|---|---|---|
| `uci_benchmark.h5` | UCI [Musk v1](https://archive.ics.uci.edu/dataset/74/musk+version+1) / [v2](https://archive.ics.uci.edu/dataset/75/musk+version+2), [Letter Recognition](https://archive.ics.uci.edu/dataset/59/letter+recognition), [HEPMASS](https://archive.ics.uci.edu/dataset/347/hepmass) | CC BY 4.0 |
| `mnist_xai_benchmark_100bags.h5` | [MNIST](http://yann.lecun.com/exdb/mnist/) | CC BY-SA 3.0 |
| `andrews_mil_benchmark.h5` | [Fox / Tiger / Elephant](https://www.cs.columbia.edu/~andrews/mil/datasets.html) image bags | none stated |
| `tcga_uni2_luad_vs_lusc.h5` | TCGA LUAD/LUSC slides, [UNI2-h](https://huggingface.co/MahmoodLab/UNI2-h) features | TCGA open access; UNI2-h CC BY-NC-ND 4.0 |
| `rsna_ich_resnet50_draws_100bags.h5` | [RSNA-ICH](https://www.kaggle.com/competitions/rsna-intracranial-hemorrhage-detection), via [torchmil](https://huggingface.co/datasets/torchmil/RSNA_ICH_MIL) | RSNA challenge terms |

## Citation

```bibtex
@article{mollers2026incontext,
  title   = {In-Context Multiple Instance Learning},
  author  = {M\"ollers, Alexander and Sextro, Marvin and Hense, Julius and Dernbach, Gabriel and M\"uller, Klaus-Robert},
  journal = {arXiv preprint arXiv:2606.06458},
  year    = {2026}
}
```