Ghana-Chat

Ghana-Chat is a large-scale, high-quality synthetic conversational dataset designed specifically to ground Large Language Models (LLMs) in the Ghanaian context.

With over 226,000 multi-turn conversations, this dataset bridges the representation gap in modern AI, providing rich, factual, and culturally relevant dialogue based entirely on real Ghanaian news articles, international news reported from a Ghanaian perspective, and academic research.

Purpose and Motivation

Mainstream LLMs often struggle with localized knowledge, hallucinating facts or providing generic answers when asked about specific African contexts.

Ghana-Chat aims to fix this by providing training data that teaches models how to discuss:

• Local Ghanaian news, politics, and current events, as well as international news reported through a Ghanaian lens.
• Ghanaian culture, economics, and history.
• Scientific and academic research specific to Ghana.

By fine-tuning on this dataset, models can learn to answer user questions about Ghana accurately, deeply, and in a natural conversational tone.

Dataset Overview

• Total Conversations: ~226,300
• Total Turns (Rows): ~1.97 Million
• Language: English
• Format: Multi-turn chat (UltraChat style)
• Sources: Local and international news articles from Ghanaian publishers, plus research papers.

How It Was Built

To ensure high factual accuracy and deep contextual grounding, this dataset was generated using Llama 3.1 70B (meta/llama-3.1-70b-instruct) in a community-driven volunteer effort led by the Ghana NLP Community.

1. Sourcing: We collected real, factual texts consisting of Ghanaian news articles and research excerpts.
2. Prompting: The LLM was provided with these texts and instructed to generate a 4-to-6 turn conversation between a curious user and a knowledgeable assistant.
3. Strict Grounding: The model was heavily constrained to base all answers strictly on the provided text, using exact dates, facts, and figures without hallucination.
4. Formatting: The output was structured into standard user/assistant conversational roles, progressing from basic inquiries to deeper, analytical questions.

Data Structure

The dataset is formatted to be completely ready for Hugging Face chat templates (like Llama 3, Mistral, or ChatML).

To ensure clean, focused training, the dataset has been stripped down to its core conversational components. Each row represents a full multi-turn conversation and contains the following key fields:

Field	Type	Description
unique_conv_id	string	A unique identifier for the specific conversation thread.
source	string	The source i.e. research or news.
messages	list[dict]	The actual conversation turns. Each item is a dictionary containing {"role": "...", "content": "..."}.

Example Message Structure

[
  {"role": "user", "content": "Can you explain the recent developments in the agricultural sector according to the news?"},
  {"role": "assistant", "content": "Based on the article, the Ministry of Food and Agriculture announced a new subsidy program on September 5th aimed at local cocoa farmers..."},
  {"role": "user", "content": "How will this subsidy impact the farmers specifically?"},
  {"role": "assistant", "content": "The subsidy reduces the cost of fertilizers by 30%, which is expected to..."}
]

Use Cases

• Supervised Fine-Tuning (SFT): Perfect for fine-tuning base models into chat assistants that are highly knowledgeable about Ghana.
• RAG Evaluation: Use the source texts and resulting conversations to evaluate Retrieval-Augmented Generation systems.
• Cultural Alignment: Aligning models to better understand the nuances of Ghanaian society and local reporting.

Contributors & Owners

This dataset is owned and was created by the following core contributors, alongside the incredible volunteers in the GhanaNLP Community who donated their time and compute to process millions of rows of data:

• Mich-Seth Owusu
• Kasuadana Adams
• Emmanuel Saah
• Alidu Abubakari
• Gerhardt Datsomor
• Kelvin Newman

Citation

If you use this dataset in your research or applications, please cite it as follows:

APA Style:

Owusu, M.-S., Adams, K., Saah, E., Abubakari, A., Datsomor, G., & Newman, K. (2026). Ghana-Chat: A Synthetic Conversational Dataset for Grounding LLMs in the Ghanaian Context. Hugging Face. https://huggingface.co/datasets/ghananlpcommunity/ghana-chat

BibTeX:

@misc{ghanachat2026,
  author = {Owusu Mich-Seth; Adams Kasuadana; Saah Emmanuel; Abubakari Alidu; Datsomor Gerhardt; Newman Kelvin},
  title = {Ghana-Chat: A Synthetic Conversational Dataset for Grounding LLMs in the Ghanaian Context},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{[https://huggingface.co/datasets/ghananlpcommunity/ghana-chat](https://huggingface.co/datasets/ghananlpcommunity/ghana-chat)}}
}

License

MIT License