nigeria-oil-spill-analysis-model

Nigeria Oil Spill Analysis Model

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 raw oil spill statistics from Nigeria's NOSDRA and NUPRC regulatory data and produce structured environmental analytical reasoning, not question-and-answer pairs. Raw statistics go in, expert interpretation comes out, grounded strictly in cited source data.

Training Data

• Source: NOSDRA via TheCable and Nairametrics, NUPRC via Platform Times and TheCable, Frontiers in Environmental Science, FairPlanet, Pulitzer Center, ScienceDirect, climatechangenews.com
• Dataset: 5 original cited prompt-completion pairs, expanded via Adaptive Data (House Special + Hallucination mitigation only, Reasoning Traces deliberately excluded)
• Languages: English, Hausa, Yoruba
• Quality improvement: 18.8% relative (Grade B → A, score 8.0 → 9.5)
• Kaggle: https://www.kaggle.com/datasets/yunusahusseinadeiza/nigeria-oil-spill-environmental-interpreter
• Hugging Face dataset: https://huggingface.co/datasets/mabera/nigeria-oil-spill-dataset

Training Metrics

• Win rate (on dataset): 73% adapted vs 27% base model
• General Win Rate (unseen Science-domain tasks): 79% adapted vs 21% base
• Base model: meta-llama/Llama-3.3-70B-Instruct
• Method: LoRA — House Special + Hallucination mitigation
• Dataset quality: 8.0 → 9.5 (+18.8% relative improvement, Grade A)

Why the General Win Rate Result Matters

At 79% on unseen Science-domain tasks, this is the highest General Win Rate in this portfolio, beating even the Desertification model's 77%. The model scored higher on unseen tasks than on its own training distribution (79% vs 73%), confirming the same healthy generalization pattern seen on the previous submission and suggesting these adapters are not narrowly overfitting on their training data.

Recipe Note

Reasoning Traces were deliberately excluded from this submission following a finding on a prior submission (Nigeria Desertification Analysis Model) where expanded training rows included elaboration not traceable to the original cited sources. House Special + Hallucination mitigation alone produced Grade A results with the highest General Win Rate in this portfolio, confirming this was the right call.

Key Cited Findings (from original source data only)

• Nigeria's oil spill incident count and spilled volume have decoupled since 2021: 2023 had roughly 3x the incidents of 2021 but a lower spilled volume, meaning neither metric alone tells the full story
• Two government agencies (NOSDRA and NUPRC) reported a 24% discrepancy in incident counts for the same year (589 vs 732 in 2024), consistent with a documented historical pattern of 300 to 3,000 occurrences divergence between agencies and operators
• No legally binding penalties or fines currently exist for oil spills in Nigeria
• Ogoniland remediation stands at approximately 11% completion 13 years into a 25 to 30 year UN-mandated timeline, with 2024 field assessments finding contamination 260x above regulatory standards at officially certified remediated sites

Credits

Powered by Adaptive Data — Adaption Labs AutoScientist Challenge 2026