k-kiirikki/trust-dvine-4d

---
datasets:
- k-kiirikki/trust-dvine-4d
annotations_creators:
- no-annotation
language:
- en
license: apache-2.0
multilinguality:
- monolingual
size_categories:
- n<1K
source_datasets:
- original
task_categories:
- other
---

# Trust-DVine-4D: Simulation-Based Inference Benchmark (4D)

This dataset contains simulation-based inference results for the **4-dimensional DVine** problem from the ["Averting A Crisis in Simulation-Based Inference"](https://arxiv.org/abs/2110.06581) paper.

## Dataset Description

The DVine (Dependency Vine) problem evaluates Bayesian inference methods on a 4-dimensional parameter space using a vine copula-based simulator. This dataset contains:

### Reference Data (`vine_reference/`)
- `coverage-vine-posterior.npy` - Coverage statistics for the vine posterior
- `coverage-vine-mst.npy` - Coverage statistics for the MST (Minimum Spanning Tree) vine
- `sbc-vine-posterior.npy` - Simulation-based calibration results for vine posterior
- `sbc-vine-mst.npy` - SBC results for MST vine

### Simulation Results (`output/{sample_size}/`)
Results organized by number of simulations (sample sizes: 1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072).

Each sample size directory contains results from multiple methods (`without-regularization/`):
- **MLP methods**: `mlp-{00-04}`, `mlp-bagging-{00-04}`, `mlp-static-{00-04}`
- **Flow-SBI methods**: `flow-sbi-{00-04}`, `flow-sbi-bagging-{00-04}`, `flow-sbi-static-{00-04}`

### File Types
- `coverage.npy` - Coverage diagnostic arrays
- `sbc.npy` - Simulation-based calibration arrays
- `diagnostic.npy` - Additional diagnostic metrics
- `losses-*.npy` - Training/validation/test losses
- `test-loss-functionals.npy` - Test loss functionals
- `posterior.pkl` - Serialized posterior samples

## Citation

If you use this dataset, please cite:

```bibtex
@article{terral2022averting,
title={Averting A Crisis in Simulation-Based Inference},
author={Terral, Thomas and others},
journal={arXiv preprint arXiv:2110.06581},
year={2022}
}
```

## Related Datasets

- [k-kiirikki/trust-dvine-6d](https://huggingface.co/datasets/k-kiirikki/trust-dvine-6d) - 6-dimensional DVine results