SAHELI: A Culturally Grounded Maternal Health Conversation Dataset

Overview

SAHELI is a large-scale synthetic dialogue dataset designed to promote research in maternal health support and culturally grounded conversational AI. It comprises over 6,000 multi-turn conversations (121,200 utterances) generated from 101 demographic profiles representing urban and semi-urban Indian contexts.

Each conversation models interactions between an AI companion and an Indian woman, structured across 10 thematic dimensions—including family involvement, maternal agency, and negotiation between traditional and biomedical care—and spans two phases of the maternal journey: Pregnancy and Postpartum. The dataset captures three graded levels of behavioral variation, High, Low, and Mid, to reflect differences in communication styles and cultural orientation.

SAHELI provides a structured benchmark for studying behavioral patterns, cultural grounding, and dialogue understanding in maternal health interactions.

Dataset Description

SAHELI is a large-scale synthetic conversational dataset designed for modeling maternal health support interactions in Indian contexts. Each instance represents a multi-turn dialogue between an AI companion and an Indian woman.

The dataset is structured across:

• 10 thematic dimensions of maternal health
• 2 interaction phases (e.g., Pregnancy, Postpartum)
• 3 behavioral ranges (e.g., High, Mid, Low)

---

Data Generation Process

The dataset is fully synthetically generated using a Large Language Model (LLM).

Structured inputs are defined in a configuration file (master.json), specifying:

• theme name, code, and definition
• score range guide
• phase concerns
• behavioral markers
• assistant guidance and response cues

Prompt templates are used to dynamically construct inputs by combining these structured variables. The LLM generates multi-turn maternal health support conversations conditioned on these inputs.

---

Dataset Structure

saheli/
│
├── data/
│   └── train-00000-of-00001.parquet   # 6,060 conversations (released dataset)
│
├── master.json                        # Structured input configuration
│
├── (conceptual structure used during generation)
│
│   ├── P001/                          # Profile directories (total: 101)
│   │   ├── BODY_OPEN/                 # Theme directories (10 per profile)
│   │   │   ├── Pregnancy/             # Phase directories (2 per theme)
│   │   │   │   ├── range_label_High.json
│   │   │   │   ├── range_label_Mid.json
│   │   │   │   └── range_label_Low.json
│   │   │   └── Postpartum/
│   │   │       ├── range_label_High.json
│   │   │       └── ...
│   │   ├── COMM_NORMS/
│   │   └── ...
│   │
│   ├── P002/
│   │   └── ...
│
└── README.md                        

## Privacy & Ethics

This is a fully synthetic dataset. No real patient conversations are included.

Demographic profiles are derived from structured inputs and do not contain personally identifiable information. The dataset is designed to reflect realistic scenarios while avoiding sensitive or harmful content.

**This dataset should NOT be used for:**

- Clinical decision-making  
- Medical diagnosis or treatment  
- Deployment in real-world healthcare systems  

---

## License

This dataset is released under the **CC-BY-4.0** license.