Mirabest Confident Eval

---
name: mirabest-confident-eval
description: Evaluates the ability of self-supervised learning models to classify radio galaxy morphologies (e.g., Fanaroff-Riley classes) using standard image classification protocols. It measures how well disentangled generative augmentations and contrastive learning pipelines capture astrophysical structure compared to traditional data augmentations. Use when the user wants to benchmark on MiraBest Confident, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2602.18923
  bibtex_key: alphonse2026variational
  confidence: high
---

# mirabest-confident-eval

> Variational views for self-supervised learning in radio astronomy — Alphonse et al. (2026) (arXiv:2602.18923, 2026)

## What this evaluates

Evaluates the ability of self-supervised learning models to classify radio galaxy morphologies (e.g., Fanaroff-Riley classes) using standard image classification protocols. It measures how well disentangled generative augmentations and contrastive learning pipelines capture astrophysical structure compared to traditional data augmentations.

## Datasets

- **MiraBest Confident** — total ?; splits: test (-1)

## Metrics

- `accuracy` **(primary)** — range: [0, 1]
  - Mean classification accuracy calculated as the proportion of correctly predicted class labels out of the total number of instances. Reported as mean ± standard deviation across multiple runs or ablation settings.

## Input / output format

**Input**: Radio galaxy images processed through a self-supervised learning pipeline (e.g., BYOL) with various augmentation strategies.

**Output**: Predicted class labels (e.g., Fanaroff-Riley morphology classes) for each input image.

## Scoring recipe

```python
def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels)
```

## Common pitfalls

- Failing to report both finetuning and linear evaluation accuracies, as they capture different aspects of representation quality and can differ significantly.
- Omitting standard deviation alongside mean accuracy, which hides variance across random seeds or augmentation ablations.
- Confusing the baseline augmentation settings with the β-VAE generative view ablations, leading to incorrect comparison rows in the results table.

## Evidence (verbatim from paper)

> This appendix presents the detailed numerical results corresponding to the ablation studies discussed in Section 8.2. Table A1 provides the mean accuracy and standard deviation for both fine-tuning and linear evaluation protocols on the MiraBest Confident dataset.

## Citation

```bibtex
@misc{alphonse2026variational,
  title={Variational views for self-supervised learning in radio astronomy},
  author={Alphonse et al. (2026)},
  year={2026},
  note={arXiv:2602.18923}
}
```

- arXiv: 2602.18923