AfriGuard-afroxlmrguard

Classifier-style safety data for training AfroXLMR-Guard, derived from israel/AfriGuard-inst by parsing the XML output field (<safety>...</safety><category>...</category><response>...</response>).

Fields

Column	Description
prompt	User prompt (English or native African language)
response	Reference reply, parsed from <response>
safe_unsafe_label	safe / unsafe, parsed from <safety>
categorical_label	Harm category from the source top_category column (even for safe rows)
refusal	non-refusal if safe, refusal if unsafe (gold responses to unsafe prompts are refusals)
id, source, config, language	Provenance metadata from AfriGuard-inst

Statistics

Split	Language	Items	Safe	Unsafe	Refusals
train	amharic	960	480	480	480
train	english	36706	18380	18326	18326
train	hausa	947	475	472	472
train	igbo	947	475	472	472
train	oromo	960	480	480	480
train	shona	960	480	480	480
train	swahili	960	480	480	480
train	twi	979	490	489	489
train	wolof	965	484	481	481
train	yoruba	947	475	472	472
train	zulu	975	489	486	486
validation	amharic	20	10	10	10
validation	english	6975	3493	3482	3482
validation	hausa	20	10	10	10
validation	igbo	20	10	10	10
validation	oromo	20	10	10	10
validation	shona	20	10	10	10
validation	swahili	20	10	10	10
validation	twi	20	10	10	10
validation	wolof	20	10	10	10
validation	yoruba	20	10	10	10
validation	zulu	20	10	10	10
train total		46306	23188	23118	23118
validation total		7175	3593	3582	3582

Usage

from datasets import load_dataset

ds = load_dataset("israel/AfriGuard-afroxlmrguard")
train, val = ds["train"], ds["validation"]