Training and experimentation framework for the CARE project — Culturally Aligned Responsible Education for adolescent Sexual and Reproductive Health (SRH) in African contexts.
CARE explores how large language models can provide adolescent SRH education that is medically responsible, culturally aligned, age-appropriate, and useful within African contexts.
care-trainer contains the training, fine-tuning, data-processing, evaluation, and experimentation workflows used to develop CARE models.
Large language models can provide accessible health information, but generic models may not adequately account for:
- African cultural and social contexts
- adolescent-specific needs
- local health terminology and communication patterns
- culturally sensitive topics
- medical safety and factuality
- age-appropriate communication
- harmful stereotypes and cultural assumptions
- appropriate refusal and escalation behaviour
CARE investigates these challenges through a multidisciplinary approach combining machine learning, alignment research, social and cultural analysis, and health sciences.
This repository focuses specifically on the model development and experimentation pipeline.
care-trainer is intended to support:
- dataset preparation and validation
- data augmentation
- supervised fine-tuning
- alignment experiments
- parameter-efficient fine-tuning
- model evaluation
- safety and refusal evaluation
- cultural-alignment evaluation
- reproducible Hugging Face experiments
- experiment tracking and configuration
- model and tokenizer preparation
- research notebooks
- reproducible training pipelines
The repository is designed so that individual experiments can be reproduced, compared, and extended without coupling the research to a single model family.
The repository follows a research pipeline approximately structured as:
Research Questions
│
▼
Data Collection / Curation
│
▼
Data Validation
│
▼
Data Augmentation
│
▼
Training / Fine-Tuning
│
▼
Alignment
│
▼
Evaluation
│
├── Medical / Health Accuracy
├── Cultural Alignment
├── Age Appropriateness
├── Safety
├── Helpfulness
└── Robustness
│
▼
Analysis
│
▼
Model / Experiment Report
The exact training strategy is an active research question and may evolve as experiments are conducted.
CARE is a multidisciplinary project involving contributors from computing, social sciences, and health sciences.
| Contributor | Focus |
|---|---|
| Michael Kirinnya Mukiibi | Alignment Training & Research Lead |
| Elton Felix Mubiru | Data & Augmentation Engineering |
| Patricia Keitetsi | Social/Cultural Alignment & Human-Centred Research |
| Scovia Katusiime | LLM/ML Operations & Reproducible Hugging Face Experimentation |
| Contributor | Focus |
|---|---|
| Luttamaguzi Mark | Health Sciences / SRH domain contribution |
| Bayiga Leah | Health Sciences / SRH domain contribution |
The multidisciplinary structure is intentional: CARE is not solely a model-training problem. Responsible SRH education requires technical, cultural, social, and health-domain perspectives.
The repository is organized around reproducible research rather than a single training script.
care-trainer/
│
├── README.md
├── LICENSE
├── CITATION.cff
├── CONTRIBUTING.md
│
├── configs/
│ ├── data/
│ ├── models/
│ ├── training/
│ └── evaluation/
│
├── data/
│ ├── README.md
│ ├── raw/
│ ├── interim/
│ ├── processed/
│ └── evaluation/
│
├── notebooks/
│ ├── 00_environment/
│ ├── 01_data/
│ ├── 02_augmentation/
│ ├── 03_training/
│ ├── 04_alignment/
│ ├── 05_evaluation/
│ └── 06_analysis/
│
├── src/
│ └── care_trainer/
│ ├── data/
│ ├── augmentation/
│ ├── training/
│ ├── alignment/
│ ├── evaluation/
│ └── utils/
│
├── scripts/
│ ├── prepare_data.py
│ ├── train.py
│ ├── evaluate.py
│ └── reproduce.py
│
├── tests/
│
└── reports/
├── experiments/
└── figures/
The repository structure is intentionally modular. Individual directories may be introduced incrementally as the research matures.
Notebooks will serve as the research interface for CARE experiments.
They are intended for:
- exploratory data analysis
- dataset inspection
- augmentation experiments
- baseline experiments
- fine-tuning experiments
- alignment experiments
- evaluation
- visualization
- qualitative analysis
- experiment reporting
Production/reusable logic should eventually live under src/, while notebooks should primarily orchestrate experiments and document research decisions.
A planned notebook sequence is:
00 — Environment & Reproducibility
01 — Dataset Exploration
02 — Data Preparation
03 — Data Augmentation
04 — Baseline Models
05 — Fine-Tuning
06 — Alignment Experiments
07 — Safety & Cultural Evaluation
08 — Comparative Evaluation
09 — Error Analysis
10 — Results & Visualization
The notebook plan will evolve alongside the CARE research methodology.
This repository may contain code and non-sensitive research datasets, but it should not be used to publish private, identifiable, confidential, or otherwise restricted health information.
Sensitive datasets should remain outside the public repository and be accessed through an appropriate controlled workflow.
Publicly redistributable datasets should retain their original licenses and attribution requirements.
Dataset documentation should record, where appropriate:
- source
- license
- provenance
- intended use
- preprocessing
- augmentation
- known limitations
- demographic/contextual considerations
CARE experiments may use publicly available pretrained language models and Hugging Face-compatible tooling.
Model selection is intentionally separated from the training framework so that different model families can be evaluated under comparable experimental conditions.
Models produced through this repository should not automatically be interpreted as clinically validated systems.
A research checkpoint is not a medical device, diagnostic system, or substitute for professional healthcare.
CARE evaluation should go beyond conventional language-model metrics.
Experiments may evaluate dimensions including:
Does the model provide medically responsible and factually appropriate information?
Does the response appropriately account for the cultural and social context in which it is being used?
Is the response suitable for the intended adolescent audience?
Does the model avoid harmful, misleading, unsafe, or inappropriate responses?
Does the model actually answer the user's question clearly and constructively?
Does model behaviour remain appropriate under variations in phrasing, ambiguity, adversarial prompts, and culturally sensitive scenarios?
Evaluation criteria and datasets should be documented alongside experiments rather than treated as hidden implementation details.
Experiments should record, where applicable:
- base model
- model revision
- tokenizer
- dataset version
- dataset configuration
- preprocessing configuration
- augmentation strategy
- training configuration
- random seed
- hardware
- software dependencies
- evaluation configuration
- experiment identifier
Where possible, experiments should be reproducible from configuration files and scripts rather than relying exclusively on notebook state.
CARE follows several principles:
Multidisciplinary Technical decisions should be informed by social, cultural, and health-science perspectives.
Culturally grounded African contexts should not be treated as a single homogeneous culture.
Safety-conscious The system should prioritize responsible health communication over maximizing answer completion.
Evidence-oriented Training and alignment decisions should be supported by measurable evaluation.
Reproducible Research workflows should be documented and repeatable.
Transparent Important datasets, assumptions, limitations, and experimental decisions should be documented whenever disclosure is appropriate.
🚧 Active research project
care-trainer is under active development. APIs, datasets, training methods, evaluation protocols, and repository structure may change as the CARE research progresses.
Results generated from early experiments should be considered research results, not validated clinical evidence.
Contributions are welcome, particularly in:
- dataset engineering
- evaluation methodology
- alignment research
- cultural-context analysis
- health-domain review
- ML engineering
- reproducibility
- documentation
Before contributing data or model outputs, contributors should ensure that they have the appropriate rights and permissions to share them.
See CONTRIBUTING.md for project contribution guidelines.
If you use CARE research, datasets, tooling, or models in academic work, please cite the project using the citation information provided in CITATION.cff.
CARE is a research project exploring culturally aligned AI-assisted SRH education.
CARE models are not a substitute for qualified medical professionals, clinical services, emergency care, or individualized medical advice.
Model outputs may contain errors. Users should seek appropriate professional healthcare guidance for medical concerns.
The software in this repository is distributed under the license specified in LICENSE.
Individual datasets, pretrained models, and third-party resources may have separate licenses and terms of use. Always review the applicable license before using or redistributing them.
CARE Framework Culturally Aligned Responsible Education for adolescent Sexual and Reproductive Health in African contexts.