sheldonmainye/Kenya-M-Pesa-Synthetic-Credit-Data

Mpesa data useful for credit risk modelling # Kenya M-Pesa Synthetic Credit Dataset — Pipeline End-to-end pipeline for generating, quality-optimising, and publishing a synthetic M-Pesa credit scoring dataset for the Kenyan context. ## Architecture ``` generate.py Simulate synthetic users from FinAccess 2021 priors ↓ data/transactions_raw.parquet (raw transaction log) data/features_labels.parquet (engineered features + default label) ↓ adapt.py Adaption Labs quality optimisation + evaluation ↓ data/features_labels_adapted.parquet (quality-graded output) ↓ publish.py Hugging Face upload with dataset card + Croissant metadata ``` ## Setup ```bash pip install -r requirements.txt export ADAPTION_API_KEY="pt_live_..." # from adaptionlabs.ai export HF_TOKEN="hf_..." # from huggingface.co export HF_REPO_ID="smainye/kenya-mpesa-credit-synthetic" ``` ## Run ```bash # Full pipeline (5,000 users) python pipeline.py # Larger dataset python pipeline.py --n-users 50000 # Estimate Adaption cost only (no charges) python pipeline.py --dry-run # Generate + publish, skip Adaption python pipeline.py --skip-adapt # Generate + adapt only, no HF upload python pipeline.py --skip-publish ``` ## Files | File | Purpose | |---|---| | `generate.py` | Persona definitions, transaction simulation, feature engineering, label assignment | | `adapt.py` | Adaption Labs SDK integration — upload, run, evaluate, download | | `publish.py` | Hugging Face upload, dataset card, Croissant metadata | | `pipeline.py` | Orchestrator | | `requirements.txt` | Python dependencies | ## Persona archetypes (FinAccess 2024 priors) | Persona | Pop. share | Txn/week | Base default | |---|---|---|---| | urban_salaried | 22% | 18 | 9% | | informal_trader | 27% | 28 | 24% | | boda_worker | 14% | 42 | 32% | | rural_smallholder | 24% | 7 | 22% | | remittance_dependent | 13% | 10 | 22% | Overall dataset default rate targets ~16.6%, anchored to F …