WAZOBIALABS/nigerian-pidgin-eval

Wazobia Labs — Nigerian Pidgin Gold-Standard Evaluation Set

Version: v0.3 — May 2026 Entries: 253 dual-verified gold-standard entries Builder: Wazobia Labs License: CC-BY-4.0 — commercial use permitted with attribution Contact: wazobialabs@gmail.com

---

What This Is

Every AI lab building for Nigerian Pidgin has the same unsolved problem: they cannot evaluate whether their model actually works.

No gold-standard benchmark exists. No culturally verified test set. No sarcasm corpus that captures Nigerian deadpan. No benchmark that distinguishes forming from contempt, hustle_fatigue from hustle_energy, a sincere compliment from its sarcastic twin delivered in the exact same words with the exact same flat tone.

This evaluation set solves that.

---

Dataset Summary

Specification	Current (v0.3)
Total entries	253 gold-standard entries
Emotion categories	16 — all represented at minimum 15 entries
Sarcasm pairs	28 complete pairs (sincere + sarcastic twin)
Health domain entries	40 entries
Female speaker representation	~45%
Inter-annotator agreement	In progress — preliminary Kappa above 0.70
Built from	Wazobia Labs Nigerian Pidgin Dataset v0.8 (550 entries)
License	CC-BY-4.0

---

The 16-Category Taxonomy

This evaluation set uses the Wazobia Labs Nigerian Pidgin emotion taxonomy — 16 categories, four of which have no equivalent in any existing NLP framework:

Category	Type	Definition
forming	Nigerian-specific ★	Deliberately performing emotional indifference as social armour — acting unbothered on purpose. A mask worn in public. AI reads this as neutral and misses everything.
hustle_fatigue	Nigerian-specific ★	The specific exhaustion of sustained economic grinding where stopping is not an option. Not burnout — the kind where you cannot afford to stop.
hustle_energy	Nigerian-specific ★	Fired-up grind motivation. Defiant optimism about economic achievement. Often includes the Lagos "before 30" cultural timeline urgency.
market_energy	Nigerian-specific ★	The sharp, alert, competitive transactional register of Lagos commerce. Not aggression — a specific deal-making emotional state.
anger	Standard	Frustration, irritation, outrage
betrayal	Standard	Feeling deceived or disrespected by someone trusted
celebration	Standard	Communal wins, hype, collective joy
contempt	Standard	Cold downward dismissal — icy, not hot like anger
craving	Standard	Intense desire for something not yet present
joy	Standard	Genuine happiness, excitement, delight
neutral	Standard	No specific emotion detectable
prayer_gratitude	Cultural	Spiritual thankfulness — the Nigerian testimony register
pride	Standard	Confidence, swagger, self-assurance
sarcasm	Critical	Meaning inversion through deadpan delivery — prosodically identical to sincere speech
shock	Standard	Surprise mixed with disbelief
suspicion	Standard	Wariness, awareness of attempted manipulation

---

What Makes This Gold-Standard

Sarcasm Pairs — The Most Dangerous Gap in Nigerian Pidgin NLP

Nigerian sarcasm is delivered deadpan — prosodically identical to sincere speech. The same words, the same flat tone, the complete opposite meaning.

"You don try well well" can be the warmest praise or the coldest insult. A model cannot distinguish them without cultural context. This evaluation set contains 28 complete sarcasm pairs: the same phrase annotated twice, once sincere and once sarcastic, with annotator notes documenting the contextual conditions that determine which reading applies.

Any model that scores well on this evaluation set has learned to handle Nigerian sarcasm. Any model that does not has a serious deployment risk.

Health Domain Coverage

40 health domain entries capture how Nigerian patients actually communicate about their health — not how they communicate in clinical settings or translated health literature.

"E don do, I just dey manage" — standard models read neutral. It means a patient has given up trying to get better and is merely surviving. That is a clinical signal that requires intervention. This evaluation set tests whether your model catches it.

Health domain entries cover: maternal health, chronic illness management, medication compliance, patient-provider communication, mental health expression, and acute symptom reporting.

Cultural Annotation Methodology

Every entry in this evaluation set was annotated by a native speaker with lived fluency across three Nigerian Pidgin registers: Warri Pidgin (the original creole form), Eastern Nigerian Pidgin, and Lagos Pidgin. Cultural authority — not institutional backing — is the methodological foundation.

The annotator is an Igbo woman born in Aba, raised in Warri, educated in Owerri, living in Lagos. This trajectory is the methodology.

---

Category Coverage

All 16 emotion categories are represented at minimum 15 entries:

Category	Entries
anger	15
betrayal	15
celebration	15
contempt	15
craving	15
forming	15
hustle_energy	15
hustle_fatigue	15
joy	15
market_energy	15
neutral	15
prayer_gratitude	15
pride	15
sarcasm	28
shock	15
suspicion	15
Total	253

---

Data Fields

Each entry contains 15 fields:

Field	Description
entry_id	Unique identifier — format WZ-T-XXXX
pidgin_text	The Nigerian Pidgin text
english_gloss	Plain English meaning
sentiment	positive / negative / neutral
emotion_category	Primary emotion from 16-category taxonomy
emotion_secondary	Secondary emotion where present
register	casual / street / proverbial
sarcasm_flag	yes / no
prosody_match	matched / contradicts / neutral
annotator_notes	Cultural context and annotation rationale
topic_domain	health / money / work / social / faith / relationship / conflict / motivation / emotion / food / greeting / life
speaker_gender	female / neutral
intensity	1–5 scale
source_type	original / overheard / social_media / adapted
date_added	ISO date

---

Why Evaluation Infrastructure Matters

When Meta, Google, Microsoft, and African AI labs build Nigerian Pidgin models, they face a fundamental problem: what does "works" mean?

Existing corpora — the BBC Pidgin corpus, WAXAL ASR data — can transcribe. But can they understand?

Can they tell that "You come early today oh" is sincere praise when said to someone who arrived unexpectedly on time — and devastating sarcasm when said to someone two hours late?

Can they score hustle_fatigue versus hustle_energy — two speakers both talking about the grind, one depleted and one fired up?

Can they read "E don do, I just dey manage" as the clinical resignation it is rather than neutral filler?

Without this evaluation set, there is no way to know. Wazobia Labs builds the evaluation infrastructure. You cannot skip the evaluation.

---

Foundation Dataset

This evaluation set is drawn from and verified against:

WAZOBIALABS/nigerian-pidgin-voice-text v0.8 — 550 entries — 16 emotion categories — CC-BY-4.0

---

Version History

Version	Entries	Date	Notes
v0.1	~100	April 2026	Initial eval set structure
v0.2	230	May 2026	Expanded to 16 categories
v0.3	253	May 2026	Built from corrected v0.8 main dataset — all categories at 15+ entries, 40 health entries

---

Roadmap

Milestone	Target
Inter-annotator agreement published	June 2026
300+ entries	July 2026
Voice-matched evaluation entries	Month 4
v1.0 public release	Q3 2026

---

Citation

@dataset{wazobia_labs_pidgin_eval_2026,
  author    = {Okoye, Stephanie Nkemjika},
  title     = {Wazobia Labs Nigerian Pidgin Gold-Standard Evaluation Set},
  year      = {2026},
  version   = {0.3.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/WAZOBIALABS/nigerian-pidgin-eval},
  license   = {CC-BY-4.0},
  note      = {First gold-standard evaluation benchmark for Nigerian Pidgin emotion classification}
}

---

Licensing and Access

Published under CC-BY-4.0 — free to use for research, academic, and commercial purposes with attribution.

Enterprise licensing with support, update guarantees, and integration documentation available separately.

Contact: wazobialabs@gmail.com

---

About Wazobia Labs

Wazobia Labs builds African language AI infrastructure that does not exist but should. We identify the specific, high-value gaps in African language data that existing datasets leave open — then build exactly those gaps with commercial licensing, production-grade quality, and the cultural specificity that real AI products need.

Not another Twitter scrape. Not another scripted studio recording. We build what they deliberately left out.

HuggingFace: WAZOBIALABS Contact: wazobialabs@gmail.com Founded: Lagos, Nigeria — 2026