AfriGuard-inst

Alpaca-style instruction-tuning data for safety-aligned fine-tuning, built from the train/validation splits of israel/AfriGuard and israel/AfriGuard-XL. Test splits are excluded and reserved for evaluation.

Construction

• AfriGuard (10 languages): each row yields TWO items — one English (prompt/response) and one native-language (prompt_translated/response_translated).
• AfriGuard-XL (34 split configs, English-only): each row yields one item. Rows with missing responses were removed.
• instruction is one of 5 paraphrased safety templates (deterministic, seed=42) that embed the full 10-category taxonomy and require an XML answer:

<safety>safe|unsafe</safety>
<category>top_category or "safe"</category>
<response>final reply</response>

Fields

Column	Description
instruction	Safety-analysis template (incl. taxonomy + XML spec)
input	User prompt (English or native language)
output	Gold XML target (safety label, category, reference response)
id	Source example id + language suffix
source	AfriGuard or AfriGuard-XL
config	Source config (language or country/region)
language	Language of input/output response
safe_unsafe	Safety label
top_category	Safety category

Statistics

Split	Source	Language	Items	Safe	Unsafe
train	AfriGuard	amharic	960	480	480
train	AfriGuard	english	9600	4808	4792
train	AfriGuard	hausa	947	475	472
train	AfriGuard	igbo	947	475	472
train	AfriGuard	oromo	960	480	480
train	AfriGuard	shona	960	480	480
train	AfriGuard	swahili	960	480	480
train	AfriGuard	twi	979	490	489
train	AfriGuard	wolof	965	484	481
train	AfriGuard	yoruba	947	475	472
train	AfriGuard	zulu	975	489	486
train	AfriGuard-XL	english	27106	13572	13534
validation	AfriGuard	amharic	20	10	10
validation	AfriGuard	english	200	100	100
validation	AfriGuard	hausa	20	10	10
validation	AfriGuard	igbo	20	10	10
validation	AfriGuard	oromo	20	10	10
validation	AfriGuard	shona	20	10	10
validation	AfriGuard	swahili	20	10	10
validation	AfriGuard	twi	20	10	10
validation	AfriGuard	wolof	20	10	10
validation	AfriGuard	yoruba	20	10	10
validation	AfriGuard	zulu	20	10	10
validation	AfriGuard-XL	english	6775	3393	3382
train total			46306	23188	23118
validation total			7175	3593	3582

Usage

from datasets import load_dataset

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