Datasets:

baobabtech
/

water-conflict-classifier-evals

version stringclasses 5 values	timestamp stringdate 2026-10-22 12:45:49 2026-10-22 12:52:49	base_model stringclasses 1 value	train_size int64 2.08k 2.08k	test_size int64 828 828	full_train_size float64 2.9k 2.9k	batch_size int64 64 64	num_epochs int64 1 3	sample_size float64 2.08k 2.08k	sampling_strategy stringclasses 2 values	test_split float64 0.29 0.29	num_iterations float64 20 20	f1_micro float64 0.79 0.84	f1_macro float64 0.73 0.78	f1_samples float64 0.41 0.44	accuracy float64 0.77 0.84	hamming_loss float64 0.06 0.09	trigger_precision float64 0.82 0.83	trigger_recall float64 0.91 0.91	trigger_f1 float64 0.87 0.87	trigger_support int64 174 174	casualty_precision float64 0.85 0.88	casualty_recall float64 0.91 0.93	casualty_f1 float64 0.88 0.9	casualty_support int64 233 233	weapon_precision float64 0.3 0.44	weapon_recall float64 0.88 0.88	weapon_f1 float64 0.45 0.59	weapon_support int64 52 52	model_repo stringclasses 1 value	dataset_repo stringclasses 1 value	notes stringclasses 1 value	training_type stringclasses 1 value	n_trials float64	search_sample_size float64	search_models float64	best_hyperparameters float64	holdout_fpr float64 0.01 0.02	holdout_fpr_real float64 0.01 0.02	holdout_fpr_synthetic float64 0 0.01	dataset_version stringclasses 1 value
sweep-e1-undersampling	2026-10-22T12:45:49.524328	BAAI/bge-small-en-v1.5	2,076	828	2,904	64	1	2,076	undersampling	0.285124	20	0.791271	0.731645	0.412842	0.771739	0.088567	0.823834	0.913793	0.866485	174	0.851406	0.909871	0.879668	233	0.300654	0.884615	0.44878	52	baobabtech/water-conflict-classifier	baobabtech/water-conflict-training-data	sweep run (not published)	standard	null	null	null	null	0.008333	0.013333	0.003333	d3.0
sweep-e2-undersampling	2026-10-22T12:45:53.301364	BAAI/bge-small-en-v1.5	2,076	828	2,904	64	2	2,076	undersampling	0.285124	20	0.838839	0.784666	0.44058	0.839372	0.064815	0.828125	0.913793	0.868852	174	0.880658	0.918455	0.89916	233	0.438095	0.884615	0.585987	52	baobabtech/water-conflict-classifier	baobabtech/water-conflict-training-data	sweep run (not published)	standard	null	null	null	null	0.013333	0.023333	0.003333	d3.0
sweep-e3-undersampling	2026-10-22T12:45:56.850815	BAAI/bge-small-en-v1.5	2,076	828	2,904	64	3	2,076	undersampling	0.285124	20	0.836978	0.781284	0.442593	0.835749	0.066023	0.828125	0.913793	0.868852	174	0.874494	0.927039	0.9	233	0.425926	0.884615	0.575	52	baobabtech/water-conflict-classifier	baobabtech/water-conflict-training-data	sweep run (not published)	standard	null	null	null	null	0.015	0.02	0.01	d3.0
sweep-e2-oversampling	2026-10-22T12:46:00.475089	BAAI/bge-small-en-v1.5	2,076	828	2,904	64	2	2,076	oversampling	0.285124	20	0.838839	0.784666	0.44058	0.839372	0.064815	0.828125	0.913793	0.868852	174	0.880658	0.918455	0.89916	233	0.438095	0.884615	0.585987	52	baobabtech/water-conflict-classifier	baobabtech/water-conflict-training-data	sweep run (not published)	standard	null	null	null	null	0.013333	0.023333	0.003333	d3.0
v3.0	2026-10-22T12:52:50.302164	BAAI/bge-small-en-v1.5	2,076	828	2,904	64	2	2,076	undersampling	0.285124	20	0.838839	0.784666	0.44058	0.839372	0.064815	0.828125	0.913793	0.868852	174	0.880658	0.918455	0.89916	233	0.438095	0.884615	0.585987	52	baobabtech/water-conflict-classifier	baobabtech/water-conflict-training-data	null	standard	null	null	null	null	0.013333	0.023333	0.003333	d3.0

Water Conflict Classifier Evaluation Metrics

Evaluation metrics tracking the performance of the Water Conflict Classifier across multiple training iterations and model configurations.

Dataset Summary

This dataset contains evaluation results from training runs of the Water Conflict Classifier, a multi-label SetFit model that identifies water-related conflict events in news headlines. Each row represents one model version with comprehensive performance metrics across three classification labels: Trigger, Casualty, and Weapon.

Related Links:

🤗 Model Collection
🐙 GitHub Repository
📦 PyPI Package
🌊 Pacific Institute Water Conflict Chronology

Dataset Structure

Fields

Field	Type	Description
`version`	string	Model version identifier (v1.0, v2.0, etc.)
`timestamp`	string	Training completion timestamp
`base_model`	string	Base embedding model used
`train_size`	int	Number of training examples
`test_size`	int	Number of test examples
`f1_micro`	float	Micro-averaged F1 score
`f1_macro`	float	Macro-averaged F1 score
`accuracy`	float	Overall accuracy
`trigger_*`	float	Precision/recall/F1 for Trigger label
`casualty_*`	float	Precision/recall/F1 for Casualty label
`weapon_*`	float	Precision/recall/F1 for Weapon label
`model_repo`	string	HuggingFace model repository

Model Versions

The dataset tracks performance across different configurations:

Base models: BAAI/bge-small-en-v1.5, sentence-transformers/all-MiniLM-L6-v2
Training strategies: undersampling for class balance
Hyperparameter variations: batch size, epochs, sample size

Usage

from datasets import load_dataset

# Load the evaluation metrics
evals = load_dataset("baobabtech/water-conflict-classifier-evals")

# Compare model versions
import pandas as pd
df = pd.DataFrame(evals['train'])
print(df[['version', 'f1_macro', 'accuracy']].sort_values('f1_macro', ascending=False))

Citation

If you use this dataset or the Water Conflict Classifier in your research, please cite:

@misc{baobab_water_conflict_classifier,
  author = {Mills, Olivier},
  title = {Water Conflict Classifier: Few-Shot Learning for Water-Related Conflict Event Detection},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/baobabtech/water-conflict-classifier}}
}

License

CC-BY-NC-4.0 (Non-commercial use only)

Contact

Olivier Mills
Website: baobabtech.ai
LinkedIn: oliviermills

Downloads last month: 65

Number of rows:

Total file size:

9.1 kB

Collection including baobabtech/water-conflict-classifier-evals

Water Conflict Classifier

Collection

Models and datasets for the classification of news and event headlines aligned with the Water Conflict Chronology by the Pacific Institute • 5 items • Updated 12 days ago