Colour Mnist Bias Eval

---
name: colour-mnist-bias-eval
description: Evaluates how well a classifier maintains performance on a biased dataset when trained on different coreset selection strategies. It probes the model's robustness to dataset bias and measures the effectiveness of data frugality techniques in mitigating bias while preserving accuracy across varying data budgets. Use when the user wants to benchmark on Colour-MNIST, or asks about evaluating this task. Reports classifier performance.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2602.19789
  bibtex_key: wilson2026datafrugality
  confidence: medium
---

# colour-mnist-bias-eval

> Stop Preaching and Start Practising Data Frugality for Responsible Development of AI — Wilson et al. (2026) (arXiv:2602.19789, 2026)

## What this evaluates

Evaluates how well a classifier maintains performance on a biased dataset when trained on different coreset selection strategies. It probes the model's robustness to dataset bias and measures the effectiveness of data frugality techniques in mitigating bias while preserving accuracy across varying data budgets.

## Datasets

- **Colour-MNIST** — total ?; splits: train (-1)

## Metrics

- `classifier performance` **(primary)** — range: percent
  - Standard classification accuracy calculated as the proportion of correctly predicted labels over the total number of test instances.

## Input / output format

**Input**: RGB images from the Colour-MNIST dataset, sampled according to specified bias strengths (fractions of majority group samples: 0.0, 0.75, 0.95, 0.99) and data budgets (100, 500, 1000, 10000, 25000, 50000).

**Output**: Predicted class labels for each input image.

## Scoring recipe

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

## Common pitfalls

- Confusing bias strength (fraction of majority group) with data budget, leading to incorrect experimental conditions.
- Assuming the 'balanced' coreset method simply downsamples to equal class sizes rather than actively rebalancing during subset selection.
- Failing to report results for both aligned (no bias) and conflicting (biased) conditions, which masks the actual mitigation gains.

## Evidence (verbatim from paper)

> Figure[9] shows the results for additional bias strengths, i.e., the fraction of samples belonging to the majority group ${0.0,0.75,0.95}$ for the following data budgets: ${100,500,1000,10000,25000,50000}. Unsurprisingly, stronger bias strength lead to larger gains from rebalancing the dataset to mitigate bias, as noted in Figure[4]. ... Performance of a classifier trained on the biased Colour-MNIST dataset is reported for the case with no bias (aligned) and with bias (conflicting) for four bias strengths (fraction of samples in the majority group): 0.0%, 75%, 95%, and 99%. Three coreset methods are shown: random (baseline), reweighted, and balanced.

## Citation

```bibtex
@misc{wilson2026datafrugality,
  title={Stop Preaching and Start Practising Data Frugality for Responsible Development of AI},
  author={Wilson et al. (2026)},
  year={2026},
  note={arXiv:2602.19789}
}
```

- arXiv: 2602.19789