Discovered Blind Spots
The model exhibited severe blind spots at the intersection of public health, localized African contexts, and advanced computational math. Specifically:
Hyper-Local Context: It hallucinates generic African stereotypes when asked about specific structural drivers of HIV in Kisumu, Kenya, completely missing documented socioeconomic factors.
Interdisciplinary Math: It fails to connect Topological Data Analysis (TDA) to epidemiological data structures, confusing persistent homology with generic graph theory.
Software Engineering: It confidently hallucinates deprecated, Windows-only R code (
memory.limit()) for Linux-based Shiny server deployments.Causal Nuance: It fails to understand causal selection bias in gender-based violence datasets, defaulting to generic "big data" platitudes.
Format Failure: As a base model, it frequently ignores the implicit QA nature of the prompt and simply autocompletes with generic essay structures ("In this essay I will discuss...").
Proposed Fine-Tuning Strategy
To fix these errors, the model requires Supervised Fine-Tuning (SFT) on a highly curated, domain-specific instruction dataset.
Dataset Composition: The dataset must include localized epidemiological reports (e.g., WHO and KEMRI data), advanced mathematical topology textbooks, and real-world R/Shiny deployment logs. Crucially, it must include "chain-of-thought" examples that force the model to differentiate correlation from causal inference.
Sourcing & Size: This dataset could be assembled by scraping MedRxiv preprints, GitHub repositories for spatial epidemiology, and African demographic health surveys (DHS). Because base models suffer from catastrophic forgetting if fine-tuned on too little data, I estimate a dataset of 15,000 to 25,000 high-quality instruction pairs would be required to successfully align the model into a specialized computational public health assistant without degrading its core language capabilities.