File size: 2,636 Bytes
53d5d30 756704d 53d5d30 756704d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | ---
license: apache-2.0
language:
- en
- ha
- yo
tags:
- legal
- nigeria
- occupational-health
- workplace-safety
- compliance
- africa
- autoscientist
base_model: gpt-oss-120b
---
# 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 |