baobabtech/water-conflict-source-data

---
license: mit
task_categories:
- text-classification
language:
- en
tags:
- water-conflict
- setfit
- multi-label
size_categories:
- 1K<n<10K
---

# Water Conflict Training Dataset

This dataset contains labeled examples for training a multi-label water conflict classifier.

## Dataset Structure

### Files

- `positives.csv`: Water conflict headlines with labels (Trigger, Casualty, Weapon)
- `negatives.csv`: Non-conflict news headlines (pre-balanced with hard negatives)

### Data Format

Both files have consistent schema:

| Column | Description |
|--------|-------------|
| Headline | News headline text |
| Basis | For positives: comma-separated labels (Trigger, Casualty, Weapon). For negatives: empty string |
| priority_sample | Boolean - For negatives: True for hard negatives (water-related peaceful news), False for ACLED. For positives: always False (not applicable) |

Note: `priority_sample` exists in both files for schema consistency but only has meaning for negatives.

### Example Rows

**Positive example:**
```
Headline,Basis
"Water reservoir attacked in region X",Casualty
```

**Negative example:**
```
Headline,Basis
"Political protest unrelated to water","
```

## Labels

- **Trigger**: Water resource as conflict trigger
- **Casualty**: Water infrastructure as casualty/target
- **Weapon**: Water as weapon/tool of conflict

## Hard Negatives & Dataset Balance

The negatives dataset is pre-balanced and training-ready, including:

1. **Hard Negatives (~120 examples, ~15-20% of negatives)**: Water-related peaceful news that teaches the model "water ≠ conflict". These prevent false positives where any water mention triggers conflict classification.
- Water infrastructure projects (peaceful development)
- Water research and technology breakthroughs
- Water conservation initiatives and conferences
- Environmental water management topics

2. **ACLED Negatives (~600 examples)**: General conflict news without water mentions. Sampled from full ACLED dataset for efficient training.

The `priority_sample` column identifies hard negatives (True) vs regular negatives (False). This balanced composition eliminates the need for complex sampling logic during training - the dataset is ready to use as-is

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("baobabtech/water-conflict-training-data")
```

## Citation

If you use this dataset, please cite the original ACLED data source.