README.md

---
license: cc-by-4.0
language:
- ar
- sw
- fr
task_categories:
- text-classification
tags:
- hate-speech
- tfgbv
- gender-based-violence
- online-harassment
pretty_name: ALTO TFGBV Gold Dataset
configs:
- config_name: all
  default: true
  data_files:
  - split: train
    path:
    - ar/train.csv
    - sw/train.csv
    - fr/train.csv
  - split: test
    path:
    - ar/test.csv
    - sw/test.csv
    - fr/test.csv
- config_name: ar
  data_files:
  - split: train
    path: ar/train.csv
  - split: test
    path: ar/test.csv
  - split: full
    path: ar/full.csv
- config_name: sw
  data_files:
  - split: train
    path: sw/train.csv
  - split: test
    path: sw/test.csv
  - split: full
    path: sw/full.csv
- config_name: fr
  data_files:
  - split: train
    path: fr/train.csv
  - split: test
    path: fr/test.csv
  - split: full
    path: fr/full.csv
---

# ALTO: African and Levantine Tech-Facilitated Gender-Based Violence Corpus

> ⚠️ **Content warning:** this dataset contains real instances of hate speech,
> harassment, threats, and other tech-facilitated gender-based violence (TFGBV).

Annotated dataset for **tech-facilitated gender-based violence (TFGBV)
classification** in **Levantine Arabic (ar)**, **Swahili (sw)**, and **African French (fr)**, collected 
from community-operated tiplines and social media. The dataset was annotated through an active-learning human-in-the-loop pipeline.

## Usage

```python
from datasets import load_dataset

alto_ds = load_dataset("meedan/alto")            # all three languages
alto_ds_sw = load_dataset("meedan/alto", "sw")   # single language: "ar", "sw", or "fr"
```

## Dataset Statistics

### Splits
The below are the split counts for the train-test split for instances with adjudicated or fully agreed (gold and silver) binary TFGBV labels.

| config | train | test | total | TFGBV+ train | TFGBV+ test |
|---|---|---|---|---|---|
| **ar** | 467 | 117 | 584 | 206 (44.1%) | 52 (44.4%) |
| **fr** | 418 | 105 | 523 | 181 (43.3%) | 46 (43.8%) |
| **sw** | 508 | 127 | 635 | 193 (38.0%) | 48 (37.8%) |
| **all** | 1393 | 349 | 1742 | 580 | 146 |

The train/test split is an 80/20 `train_test_split` stratified jointly on `tier1` × `batch` (`random_state=42`; a stratum with a single member is merged into the largest stratum of the same `tier1` before splitting), so every annotation round and every tier-1 category is represented proportionally in both splits.

### Annotation quality by language

| language | gold_agreement | gold_adjudicated | silver | bronze | total |
|---|---|---|---|---|---|
| **ar** | 404 | 46 | 279 | 217 | 946 |
| **fr** | 297 | 85 | 182 | 83 | 647 |
| **sw** | 364 | 115 | 273 | 284 | 1036 |


## Schema

| column             | type   | description |
|--------------------|--------|-------------|
| `text`             | string | Original post text |
| `english_text`     | string | English translation of the text |
| `simple_hate`      | int    | Binary hate/toxicity label (0/1) - hard label |
| `gendered_content` | int    | Whether the content is gendered (0/1) - hard label |
| `tier1`            | string | Tier-1 TFGBV taxonomy category hard label (`non_tfgbv`, `harassment_and_hate_speech`, `threats_and_incitement_t_i_of_harm_and_violence`, `image_based_abuse`, `doxxing`) |
| `tier2`            | list(string) | Tier-2 subcategories hard labels (up to 2) within the Tier-1 category |
| `tfgbv`            | int    | Binary TFGBV hard label (0/1) derived from `gendered_content==1` and `tier1 != non_tfgbv`, null when no agreement or adjudication|
| `batch`            | int    | Active-learning annotation round (1–6) the item was labelled in |
| `status` | string | Annotation status: `gold` (adjudicated or full agreement), `silver` (Missing one annotation target agreement from `gendered_content`, `tier1`, or `tier2`) |
| `status_detail` | string | Indicates the detail of the status for gold whether adjudicated or full agreement |
| `language`         | string | `ar`, `sw`, or `fr` |
|`annotators` | list(string) | Anonymized codes of the annotators who labelled the item |
|`soft_gendered` | list(string) | Per-annotator gendered answer, aligned with `annotators` |
|`soft_tier1` | list(string) | Per-annotator Tier-1 label, aligned with `annotators` |
|`soft_tier2` | list(string) | Per-annotator Tier-2 labels, aligned with `annotators` |


## Data Sources

- **Sources:** Civil society partner-operated tiplines and social media collection based on keywords.
- **Annotation:** multi-annotator labelling in Label Studio against a two-tier TFGBV
  taxonomy; items adjudicated by an expert annotator.
- **Sampling:** annotation batches were selected via an active-learning loop
  (diversity/uncertainty sampling) rather than at random, so label distributions do not
  reflect base rates in the wild.

Dataset produced by [Meedan](https://meedan.com) and collaborators.