nigeria-desertification-analysis-model

Nigeria Desertification 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

This is a LoRA adapter fine-tuned to interpret raw desertification statistics from Nigeria's 11 frontline Sahel-bordering states and produce structured environmental analytical reasoning, grounded in peer-reviewed remote sensing research and UNCCD/UNEP figures.

⚠️ Known Limitation — Please Read Before Use

During Adaptive Data expansion (House Special + Reasoning Traces), the generated training rows included elaboration beyond what the 5 original cited source rows actually support, specific tree species, a numeric "forest half-life" estimate, and an albedo feedback-loop explanation that do not trace to any cited source. This was caught at the recipe-review stage but the dataset was adapted with Reasoning Traces enabled anyway.

Treat any specific figure or mechanism from this model that is not also present in the original 5-row source dataset as unverified model elaboration, not a cited fact. The original key-findings reference table (linked below) reflects only genuinely cited statistics.

Update: raised this directly with the Adaption team at the June 25 Research Hour. They confirmed this is a known gap and that stricter source-grounding for Reasoning Traces expansion is being addressed on their end.

Training Data

• Source: UNCCD Nigeria Country Profile, UNEP, Sambe et al. 2026, Ibrahim et al. 2022
• Dataset: 5 original cited prompt-completion pairs, expanded via Adaptive Data (see limitation above)
• Languages: English, Hausa, Yoruba
• Quality improvement: 31.4% (Grade B → A)
• Kaggle: https://www.kaggle.com/datasets/yunusahusseinadeiza/nigeria-desertification-environmental-interpreter
• Hugging Face dataset: https://huggingface.co/datasets/mabera/nigeria-desertification-dataset

Training Metrics

• Win rate (on dataset): 73% adapted vs 27% base model
• General Win Rate (unseen Science-domain tasks): 77% adapted vs 23% base
• Base model: meta-llama/Llama-3.3-70B-Instruct
• Method: LoRA — House Special + Reasoning Traces + Hallucination mitigation
• Dataset quality: 7.0 → 9.2 (+31.4% improvement, Grade A)

Why the General Win Rate Result Matters

This is the first submission in my portfolio evaluated against Adaption's new global held-out test set rather than only the training-specific metric. The model scored higher on unseen Science-domain tasks (77%) than on its own training distribution (73%), a positive signal against overfitting on this small 5-row source dataset.

Key Cited Findings (from original source data only)

• Nigeria's frontline states saw 14x more deforestation than successful afforestation over a 25-year remote sensing study
• Desertification continued expanding even in years with more favorable rainfall and temperature, suggesting drivers are substantially decoupled from climate variability
• Overgrazing accounts for ~58% of land degradation in the region

Credits

Powered by Adaptive Data — Adaption Labs AutoScientist Challenge 2026