nigeria-amr-classifier-v2

Nigeria AMR Closed-Label Classifier v2

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 for closed-label antimicrobial resistance classification, combining individual MIC-based susceptibility classification (Susceptible/Intermediate/Resistant per CLSI M100) with population-level resistance rate classification (Low/Moderate/ High) from real Nigerian national surveillance data. Unlike open-ended interpretation, every classification is deterministically verifiable against a fixed ground truth.

Methodology: Why This Is a Closed-Label Task

This dataset was deliberately trained with no recipe modifications (no House Special, no Hallucination Mitigation, no Reasoning Traces). This follows methodology pioneered by other AutoScientist Challenge participants working on closed-label classification tasks, where generation recipes designed for open-ended prose can introduce label drift outside a fixed taxonomy. A system-level constraint was applied instead, restricting output to the exact valid labels only.

Ablation Test: Confirming This Choice With First-Party Evidence

Rather than rely solely on other builders' reported findings, a controlled ablation experiment was run on this exact dataset (identical 34 rows, identical expansion settings) with House Special enabled, to directly test whether it helps or hurts this closed-label task.

Enabling House Special measurably hurt every metric on this task. This confirms, with first-party controlled evidence rather than borrowed reasoning, that the "no recipe modifications" choice was correct for this closed-label classification dataset. Ablation dataset ID: c2f6e743-ca4e-4da7-9dbd-3d565229fcb0 (not published as a standalone model, since it exists solely as a negative control for this comparison).

Training Data

• Source: CLSI M100 (35th Edition, 2025); MAAP/Fleming Fund Nigeria Country Report (2022), covering 25 sentinel laboratories and 23,963 positive cultures (2016-2018); PLoS One; BMC Infectious Diseases
• Dataset: 34 rows total — 6 individual MIC classifications + 28 national surveillance rate classifications, expanded via Adaptive Data (no recipe modifications, minimum general-purpose diversity only)
• Kaggle: https://www.kaggle.com/datasets/yunusahusseinadeiza/amr-classifier-v2-national-surveillance

Verification

Every classification in the source dataset was independently, programmatically verified before training:

• 6 MIC classifications checked against the CLSI breakpoint registry
• 28 surveillance rates recomputed from raw N/n counts and matched exactly against the national report's own published percentages
• 100% pass rate, demonstrated live in the accompanying Kaggle notebook

Training Metrics

• Win rate (on dataset): 58% adapted vs 42% base model
• General Win Rate (Science-domain tasks, correctly classified): 71% adapted vs 29% base
• Base model: meta-llama/Llama-3.3-70B-Instruct
• Method: LoRA — no recipe modifications, system-level closed-label constraint
• Dataset quality: 8.0 → 8.7 (+8.7% relative improvement, Grade B)
• Percentile: 31.5 (highest baseline percentile in this author's portfolio)

Key Cited Findings (from verified source data only)

• 3rd-generation cephalosporin resistance in Enterobacterales: 67-73% across 2016-2018, among the highest rates recorded nationally
• MRSA rates in Nigeria reached 81.6% by 2018, among the highest documented globally
• Carbapenem resistance in Enterobacterales fell from 19.0% (2016) to 14.5% (2018), one of the few improving trends in the dataset
• Resistant infections carry an 84% higher mortality risk globally; MDR bloodstream infections showed 32.1% 30-day mortality versus 18.8% for non-MDR infections in a 2022-2024 African cohort study

Credits

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