README.md

---
datasets:
- LabHC/bias_in_bios
language:
- en
base_model:
- FacebookAI/roberta-base
pipeline_tag: text-classification
---

# RoBERTa-Bios-biased

This model is a `roberta-base` model fine-tuned for profession classification on a modified version of the [`LabHC/bias_in_bios`](https://huggingface.co/datasets/LabHC/bias_in_bios) dataset.

It was trained to study the impact of amplified gender/profession correlations in the training data. Compared with `roberta-bios`, this model was trained on a biased version of the BIOS training split.

## Model details

* Base model: `roberta-base`
* Dataset: `LabHC/bias_in_bios`
* Input column: `hard_text`
* Label column: `profession`
* Gender column used to modify the training set: `gender`
* Task: profession classification
* Language: English


## Biased training data construction

The model was trained on a modified version of the BIOS training split.

For each profession, the gender distribution was computed. If one gender represented more than 65% of the examples for a given profession, this profession was considered biased. For these professions, only examples from the majority gender were kept. For professions without a majority gender above this threshold, all examples were kept.

The threshold used was:

```python
THRESHOLD = 0.65
```

In simplified form:

```python
if majority_gender_ratio > 0.65:
    keep only examples from the majority gender for this profession
else:
    keep all examples for this profession
```

This procedure deliberately amplifies gender/profession correlations in the training data.

## Training procedure

The model was fine-tuned with the Hugging Face `Trainer` API.

Main hyperparameters:

```python
BASE_MODEL = "roberta-base"
MAX_LENGTH = 256
NUM_EPOCHS = 3
LEARNING_RATE = 2e-5
TRAIN_BATCH_SIZE = 32
EVAL_BATCH_SIZE = 128
SEED = 42
```

The model was trained using:

```python
AutoModelForSequenceClassification.from_pretrained(
    "roberta-base",
    num_labels=num_labels,
)
```

The best checkpoint was selected according to macro-F1 on the development split.

## Evaluation

Reported performance:

| Evaluation set                   | Accuracy |
| -------------------------------- | -------: |
| Modified BIOS test set |   0.8779 |
| Original BIOS test set           |   0.8539 |