---
base_model: Qwen/Qwen3-8B
library_name: peft
license: mit
pipeline_tag: text-generation
tags:
- base_model:Qwen/Qwen3-8B
- llama-factory
- lora
- transformers
- medical
- public-health
model-index:
- name: sft100
  results: []
---

# GlobalHealthAtlas Public Model

This repository contains the fine-tuned LoRA adapter for the **GlobalHealthAtlas Public Model**, introduced in the paper [From Knowledge to Inference: Formalizing Specialized Public Health Reasoning on GlobalHealthAtlas](https://huggingface.co/papers/2602.00491).

The model is designed to provide informative, context-aware answers to public health–related queries across 15 domains and 17 languages. It was fine-tuned from [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) on the GlobalHealthAtlas dataset.

- **Paper:** [From Knowledge to Inference: Formalizing Specialized Public Health Reasoning on GlobalHealthAtlas](https://huggingface.co/papers/2602.00491)
- **Repository:** [Jan8217/GlobalHealthAtlas](https://github.com/Jan8217/GlobalHealthAtlas)

## Overview

Public health reasoning requires population-level inference grounded in scientific evidence, expert consensus, and safety constraints. This model addresses the scarcity of supervised signals in this domain. It was trained on the GlobalHealthAtlas dataset, a large-scale multilingual corpus of 280,210 instances spanning 15 domains including infectious disease prevention, health policy, and vaccination.

## Intended Uses & Limitations

This model is intended to be used as a question-answering and information retrieval component for research and public health queries. 

**Important Disclaimer:** This model is **NOT** intended for clinical diagnosis, medical advice, or other high‑stakes decision-making without human review by domain experts. 

## Training Procedure

The model was fine-tuned using LoRA (Low-Rank Adaptation) on the Qwen3-8B base model via the [LLaMA Factory](https://github.com/hiyouga/LLaMA-Factory) framework.

### Training Hyperparameters

The following hyperparameters were used during training:
- **Learning rate:** 5e-05
- **Train batch size:** 1
- **Gradient accumulation steps:** 8
- **Total train batch size:** 8
- **Optimizer:** AdamW
- **LR scheduler type:** cosine
- **LR scheduler warmup ratio:** 0.1
- **Num epochs:** 2.0
- **Seed:** 42

### Framework Versions

- PEFT 0.15.1
- Transformers 4.51.3
- Pytorch 2.3.0+cu121
- Datasets 3.2.0
- Tokenizers 0.21.0

## Citation

If you use this model or the GlobalHealthAtlas dataset, please cite:

```bibtex
@article{globalhealthatlas2026,
  title={From Knowledge to Inference: Formalizing Specialized Public Health Reasoning on GlobalHealthAtlas},
  author={GlobalHealthAtlas Team},
  year={2026}
}
```