Spaces:
Running
Running
| """Nigerian style layer — bonus marks via cultural contextualization. | |
| The challenge brief awards extra credit for systems that *behave and sound | |
| like Nigerians*. We treat this as a toggleable rendering layer, not as a | |
| core dependency, for two reasons: | |
| 1. Eval datasets are English Amazon reviews. Rendering everything in | |
| Nigerian register would hurt our ROUGE / BERTScore against the | |
| ground truth. | |
| 2. Keeping it as a flag means we can showcase the capability without | |
| sacrificing benchmark scores. Best of both rubric worlds. | |
| Two functions: | |
| naija_style_review(text) → rewrites a generated review in Nigerian | |
| English register, preserving sentiment, | |
| rating intent, and key entities. | |
| naija_persona_examples() → returns hand-crafted Nigerian personas the | |
| judges can demo Task B against. These show | |
| the system handling local taste profiles | |
| (afrobeats, jollof, Nollywood, etc.) even | |
| when the underlying catalog is Amazon-global. | |
| Design note: the style layer renders output in rich, expressive Nigerian | |
| Pidgin — confident and fluent across the whole text, the way a Nigerian | |
| genuinely talks when giving a strong opinion. Sentiment, rating intent and | |
| factual content are always preserved; only the register changes. | |
| """ | |
| from __future__ import annotations | |
| from core.llm import LLMClient | |
| NAIJA_STYLE_SYSTEM = """You are a stylist who rewrites text in rich, expressive Nigerian Pidgin English — the way a Nigerian would genuinely talk when sharing strong opinions. Rules: | |
| - Keep the sentiment, rating intent, and all factual content unchanged. A positive review stays positive; a 2-star pan stays a pan; named items, authors, and plot facts stay accurate. | |
| - Write FULLY in Nigerian Pidgin register — not standard English with a sprinkle. Lean into it confidently across the whole text. | |
| - Use natural Pidgin grammar and vocabulary throughout. Examples of the texture wanted: | |
| · "This book sweet me die, I no fit drop am at all." | |
| · "Abeg, the storyline just dey drag, e tire me well well." | |
| · "Na correct work be this — the writer sabi wetin e dey do." | |
| · "I no go lie, the ending shock me, I no see am coming." | |
| · "The characters dey alive, you go feel like say you sabi them." | |
| · "E no make sense, I vex small as I read am finish." | |
| · "This one na better book, e make sense gan-gan." | |
| - Common markers to use freely: "abeg", "sha", "na", "dey", "wetin", "e be like say", "no be small thing", "gan-gan", "well well", "I no go lie", "comot", "sabi", "vex", "sweet me", "make sense". | |
| - Keep it authentic, not caricature — write like a real Nigerian sharing a genuine opinion, not a parody. It should read as natural Pidgin, fluent and confident. | |
| - Do NOT add cultural references that weren't in the original (no jollof, Lagos traffic, etc. unless the source mentioned them). | |
| - Length should stay roughly the same. | |
| - Return ONLY the rewritten text. No preamble, no explanation.""" | |
| def naija_style_review(text: str, llm: LLMClient | None = None) -> str: | |
| """Rewrite an English review in Nigerian English register. | |
| Idempotent on already-Naija text in practice (the model leaves natural | |
| phrasings alone). | |
| """ | |
| llm = llm or LLMClient() | |
| return llm.complete( | |
| prompt=f"Rewrite this review in Nigerian English register:\n\n{text}", | |
| system=NAIJA_STYLE_SYSTEM, | |
| model="bulk", | |
| ).strip() | |
| # ────────────────────────────────────────────────────────────────────────────── | |
| # Demo personas — used in the Streamlit UI to showcase cold-start handling | |
| # ────────────────────────────────────────────────────────────────────────────── | |
| NAIJA_DEMO_PERSONAS: list[dict] = [ | |
| { | |
| "name": "Tunde — Lagos software engineer", | |
| "description": ( | |
| "A 28-year-old software engineer in Lagos who reads mostly non-fiction " | |
| "(business biographies, productivity, AI/tech), watches African and " | |
| "international thrillers, and complains when books are padded or movies " | |
| "are too slow. Prefers concise, practical writing. Gives 5 stars only " | |
| "when something genuinely changed his thinking; defaults to 4. " | |
| "Frequently mentions 'value for time' and 'execution'." | |
| ), | |
| "stated_preferences": ["business biographies", "AI and tech books", | |
| "fast-paced thrillers", "Nollywood crime dramas", | |
| "concise practical writing"], | |
| "deal_breakers": ["padded chapters", "slow pacing", "academic jargon"], | |
| }, | |
| { | |
| "name": "Ngozi — Abuja public health doctor", | |
| "description": ( | |
| "A 35-year-old doctor in Abuja who reads literary fiction and African " | |
| "memoirs, watches character-driven dramas (West African and global), " | |
| "and dislikes anything that handles women's lives shallowly. Writes " | |
| "thoughtful, longer-than-average reviews. Rates with a tough 3.5 average. " | |
| "Often references 'emotional truth' and 'craft'." | |
| ), | |
| "stated_preferences": ["literary fiction", "African memoirs", | |
| "character-driven dramas", "Adichie-adjacent voice"], | |
| "deal_breakers": ["shallow female characters", "trauma porn", | |
| "lazy plotting"], | |
| }, | |
| { | |
| "name": "Bayo — Ibadan undergraduate", | |
| "description": ( | |
| "A 21-year-old student in Ibadan who reads YA fantasy, plays a lot of " | |
| "Afrobeats during study sessions, watches anime and Nollywood comedies, " | |
| "and writes short bursty reviews. Quick to give 5 stars when entertained. " | |
| "Mentions vibes, pacing, and whether something 'hits'." | |
| ), | |
| "stated_preferences": ["YA fantasy", "anime", "Nollywood comedies", | |
| "fast-paced action"], | |
| "deal_breakers": ["long descriptive passages", "overly serious tone"], | |
| }, | |
| ] | |
| def naija_persona_examples() -> list[dict]: | |
| """Return demo personas for the Task B UI's cold-start showcase.""" | |
| return [dict(p) for p in NAIJA_DEMO_PERSONAS] | |