mabera/nigeria-energy-access-interpreter

Nigeria Energy Access & Generation Mix Interpreter

Author: Hussein Adeiza (mabera) Role: Licensed Environmental Health Officer, Abuja Nigeria Base Model: Llama 3.3 70B Fine-tuned with: AutoScientist by Adaption Labs

Model Description

A LoRA adapter fine-tuned to interpret Nigeria's electricity generation and access data, producing structured analytical reasoning grounded entirely in real, directly-downloaded statistics. Raw statistics go in, expert quantitative interpretation comes out.

What Makes This Dataset Different

Unlike this author's other submissions, which cited real figures found through research, every number in this dataset's training data was pulled programmatically from a raw, unmodified source file, Our World in Data's official energy dataset, downloaded directly from github.com/owid/energy-data. No statistic was hand-typed. Every claim in the training data is independently re-verifiable against the raw 9.2MB source file included in the dataset repository.

Training Data

• Source: Our World in Data, owid/energy-data (GitHub), downloaded directly, 2025 release
• Dataset: 5 original prompt-completion pairs, every number computed via pandas directly from the raw source file, expanded via Adaptive Data (Hallucination Mitigation, full 20K+ datapoint expansion)
• Kaggle: https://www.kaggle.com/datasets/yunusahusseinadeiza/nigeria-energy-access-interpreter

Training Metrics — Strongest Result in This Portfolio

• Win rate (on dataset): 88% adapted vs 12% base model
• General Win Rate (Data-analysis-visualization domain): 81% adapted vs 19% base
• Base model: meta-llama/Llama-3.3-70B-Instruct
• Method: LoRA — Hallucination Mitigation, full 20,000+ datapoint expansion
• Dataset quality: 8.0 → 8.8 (+10.0% relative improvement, Grade B)
• Percentile: 31.5
• Domain classification: Data-analysis-visualization (60%) / Science (40%), a clean, accurate match to the intended category

Why This Result Stands Out

This is the strongest win rate and General Win Rate in this author's entire AutoScientist Challenge portfolio (15 submissions across Parts 1 and 2), exceeding even the closed-label AMR and Oil Spill submissions. The likely contributing factors: genuinely verifiable ground-truth data sourced directly from a raw file rather than research synthesis, a clean domain match routing to the strong base model, and the full datapoint expansion (unlike the closed-label submissions, this dataset carries no risk of label-taxonomy drift, so the fuller expansion could be used without the precision concerns that applied elsewhere).

Key Cited Findings (from the raw downloaded source only)

• Nigeria's per-capita electricity access (174.9 kWh/year) sits at roughly 1/22nd of the world average (3,859.8 kWh) and South Africa's level (3,749.3 kWh), despite Nigeria's far larger population (237.5 million vs South Africa's 64.7 million)
• Nigeria's generation mix shifted meaningfully in a single year: gas share fell from 75.0% (2024) to 68.7% (2025), while hydro rose from 24.5% to 30.9%
• Over the full 25-year period (2000-2025), the long-term trend remains toward increasing gas dependency (+6.9 percentage points), despite the 2025 single-year uptick in hydro
• Nigeria's solar share of generation (0.313%) sits at roughly 3.6% of the world average (8.745%), despite Nigeria's equatorial location giving it substantially higher solar potential than most solar- adopting nations

Credits

Powered by Adaptive Data — Adaption Labs AutoScientist Challenge 2026, Part 2 — Data Visualization Category