shmuhammadd/ghana-masterclass-synthetic-data

# AI and Synthetic Data in Low-Resource Settings **The Ghana Scholars and Practitioners' Research Masterclass — Imperial Global Ghana** Shamsuddeen Hassan Muhammad · Imperial College London · Accra, 9 July 2026 Why the data we need often does not exist — and how generative AI can help us create it. ## Contents - `Master_class_Ghana_all.pptx` — full slide deck (59 slides) - `synthetic_data_practical.ipynb` — practical 1: synthetic **text** (sentiment) data in a Ghanaian language - `accra_household_survey_synthetic.ipynb` — practical 2: synthetic **tabular** household-survey data for Accra ## Practical 1 · Synthetic text data Build the full six-step synthetic-data pipeline from the lecture: 1. **Define** — task, language, label set, target size, schema 2. **Seed** — anchor generation in real AfriSenti examples 3. **Generate** — prompt an LLM in batches, logging every prompt 4. **Filter** — de-duplicate, drop malformed / off-label / code-switched output 5. **Review** — native speakers rate a sample 6. **Evaluate** — train a classifier on synthetic data, test on **real** data Bring a laptop — everything runs in the browser via Google Colab. Working in pairs is encouraged; non-technical participants lead the review step. > Synthetic data is a bridge, not a replacement. The final exam is always real data. ## Practical 2 · Synthetic tabular survey data Build synthetic household-survey records for Greater Accra that mirror the statistics of a real survey without describing any real family: 1. **Define** — the schema: fields, types, allowed values (the codebook) 2. **Seed** — a small survey sample whose statistics anchor generation 3. **Generate** — two routes: LLM prompting *and* a statistical generator 4. **Filter** — schema validation, range checks, dedup 5. **Review** — statistician's sanity checks on cross-field relationships 6. **Evaluate** — fidelity (distribution plots), utility (train real vs synthetic), privacy (leakage check) ## Contact Shamsudde …