Bias In Bios Eval

---
name: bias-in-bios-eval
description: This benchmark evaluates occupation prediction models for fairness regarding gender bias. It tests whether classifiers maintain consistent predictions when gender pronouns are swapped (individual fairness) and whether true positive rates are balanced between male and female bios across occupations (group fairness). Use when the user wants to benchmark on Bias in Bios, or asks about evaluating this task. Reports Balanced Accuracy (BA).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2006.14168
  bibtex_key: yurochkin2020sensi
  confidence: high
---

# bias-in-bios-eval

> SenSeI: Sensitive Set Invariance for Enforcing Individual Fairness — Yurochkin et al. (2020) (arXiv:2006.14168, 2020)

## What this evaluates

This benchmark evaluates occupation prediction models for fairness regarding gender bias. It tests whether classifiers maintain consistent predictions when gender pronouns are swapped (individual fairness) and whether true positive rates are balanced between male and female bios across occupations (group fairness).

## Datasets

- **Bias in Bios** — total 400000; splits: train (-1), test (-1)

## Metrics

- `Balanced Accuracy (BA)` **(primary)** — range: [0, 1]
  - Average of sensitivity and specificity, used to handle class imbalance across 28 occupations.
- `Counterfactual Token Fairness (CTF)` — range: [0, 1]
  - Variance across counterfactuals (gender-swapped bios) of the predicted probability.
- `Prediction Consistency (PC)` — range: [0, 1]
  - Proportion of test bios where the predicted occupation remains identical after gender pronoun swapping.
- `Gap RMS` — range: [0, 1]
  - Root mean squared gap between male and female true positive rates across 28 occupations.
- `Gap ABS` — range: [0, 1]
  - Mean absolute gap between male and female true positive rates across 28 occupations.

## Input / output format

**Input**: BERT fine-tuned feature representations of professional bio descriptions.

**Output**: Multiclass occupation prediction (28 classes).

## Scoring recipe

```python
For each test bio, swap male/female pronouns to create counterfactual.
ctf = variance(predicted_probabilities)
pc = mean(predictions == original_predictions)
gap_rms = sqrt(mean((tpr_male - tpr_female)**2))
gap_abs = mean(abs(tpr_male - tpr_female))
```

## Common pitfalls

- Counterfactuals only swap pronouns, ignoring other gendered language or structural cues.
- Group fairness is measured only on TPR gaps, not FPR or overall accuracy parity.

## Evidence (verbatim from paper)

> To compare group fairness we report root mean squared gap (Gap RMS) and mean absolute gap (Gap ABS) between male and female true positive rates for each of the occupations following prior studies of this dataset

## Citation

```bibtex
@misc{yurochkin2020sensi,
  title={SenSeI: Sensitive Set Invariance for Enforcing Individual Fairness},
  author={Yurochkin et al. (2020)},
  year={2020},
  note={arXiv:2006.14168}
}
```

- arXiv: 2006.14168