africa-hiv-funding-analysis-model

Africa HIV Funding Analysis Model

Author: Hussein Adeiza (mabera) Role: Licensed Environmental Health Officer, Abuja Nigeria Base Model: Llama 4 Scout 17B-16E Fine-tuned with: AutoScientist by Adaption Labs

Model Description

A LoRA adapter fine-tuned to interpret raw statistics on Africa's HIV funding structure and the 2025 disruption, producing structured public health analytical reasoning. Raw statistics go in, expert interpretation comes out, grounded strictly in cited source data.

Important Disclosure: Domain Classification and Model Selection

This submission was intended for the Science category, addressing genuinely epidemiological and public health content. However, the platform's automatic content classifier repeatedly labeled this dataset as News/Governance/Medical rather than Science across three separate description edits, and this classification appears to have determined which base model was used for training.

Every other submission in this author's portfolio (11 prior models) was trained on Llama 3.3 70B. This submission trained on Llama 4 Scout 17B-16E, a different and smaller model, which the author believes is a direct consequence of the News/Medical domain classification rather than a deliberate choice. This was reported to the Adaption team as a potential platform issue (bugs channel, 12 Jul 2026) prior to this model's publication.

Practical implication: this model's quality score and General Win Rate should be read in the context of a different base model and comparison pool (News Win Rates) than the author's other Science and Healthcare submissions, which were benchmarked against Llama 3.3 70B and their respective correct domains. This is not a like-for-like comparison with the rest of the portfolio.

Training Data

• Source: UNAIDS, IAS 2025 conference findings (aidsmap), Health Policy Watch, CNBC Africa, Healio
• Dataset: 5 original cited prompt-completion pairs, expanded via Adaptive Data (House Special + Hallucination mitigation)
• Languages: English, French, Portuguese
• Kaggle: https://www.kaggle.com/datasets/yunusahusseinadeiza/africa-hiv-funding-disruption-interpreter

Training Metrics

• Win rate (on dataset): 71% adapted vs 29% base model
• General Win Rate (News-domain tasks, not Science): 60% adapted vs 40% base
• Base model: meta-llama/Llama-4-Scout-17B-16E (see disclosure above)
• Method: LoRA — House Special + Hallucination mitigation, full 20K+ datapoint expansion
• Dataset quality: 6.0 → 6.5 (+8.3% relative improvement, Grade C → C)

Key Cited Findings (from original source data only)

• Donor dependency varies enormously by region and country: ~90% of ARV costs in West/Central Africa vs ~38% in East/Southern Africa, with individual countries ranging from below 25% (South Africa, Botswana, Kenya, Namibia) to above 80% (Malawi, Zimbabwe, Mozambique) for prevention funding specifically
• Nigeria's monthly PrEP initiations fell 85% (40,000 to 6,000) after 2025 budget cuts, far steeper than the 38% multi-country average decline
• Mozambique case study: 15,000+ fewer treatment starts, with children consistently harder hit than adults across every metric measured
• Modeled 2030 scenarios show a disproportionate 6-fold jump in worst-case infections between moderate cuts and full PEPFAR discontinuation, indicating treatment funding protects against new infections, not just prevention funding

Credits

Powered by Adaptive Data — Adaption Labs AutoScientist Challenge 2026, Part 2 — Science Category (see domain classification disclosure above)