Nigeria Occupational Health & Safety Law Model
AutoScientist Challenge 2026 | Legal Category
Author: Hussein Adeiza (mabera)
Role: Licensed Environmental Health Officer, Abuja Nigeria
Base Model: GPT-OSS 120B
Fine-tuned with: AutoScientist by Adaption Labs
Model Description
This is a LoRA adapter fine-tuned on Nigeria's occupational health and safety legal framework, covering the Factories Act, the Employees' Compensation Act, NSITF compliance, enforcement mechanisms and sector-specific regulations for mining, oil and gas, and radiation safety.
Training Data
- Source: Factories Act (CAP F1, LFN 2004), Employees' Compensation Act 2010, National Policy on Occupational Safety and Health 2006, ILO Nigeria Country Profile on OSH 2016
- Dataset: 10 Q&A pairs expanded via Adaptive Data
- Languages: English, Hausa, Yoruba
- Quality improvement: 12.9% (Grade C → B)
- Kaggle: https://www.kaggle.com/datasets/yunusahusseinadeiza/nigeria-occupational-health-safety-law-dataset
Training Metrics
- Win rate: 49% adapted vs 51% base model
- Base model: gpt-oss-120b
- Method: LoRA — House Special + Reasoning Traces + Hallucination mitigation
- Dataset quality: 7.0 → 7.9 (+12.9% improvement, Grade B)
Honest Finding
Unlike most of my other submissions in this challenge, this adapter did not clearly outperform its base model, gpt-oss-120b, on the official training win rate, landing at 49% adapted versus 51% base, essentially a coin flip. A likely contributing factor is the dataset's small size (10 rows) and the sharp imbalance between short prompts (11 words average) and long, detailed completions (93.7 words average), which may have given the adapter too little signal relative to an already capable 120B parameter base model on general legal reasoning. Reporting this transparently rather than only highlighting stronger results elsewhere in this challenge.
Why This Domain Matters
Nigeria's occupational safety framework spans multiple statutes, yet accessible legal guidance for employers and workers remains limited. This dataset grounds AI guidance in the actual legislative text and a cited real-world statistic (Umeokafor 2002-2021: 80% of accidents occurred at night, 53.8% of deaths concentrated in rubber manufacturing) rather than general assumptions about workplace safety.
Credits
Powered by Adaptive Data — Adaption Labs
AutoScientist Challenge 2026 | Legal Category