---
license: cc-by-nc-4.0
language:
- en
- tw
task_categories:
- question-answering
- text-generation
source_datasets:
- ghananlpcommunity/twi-english-agric
tags:
- agriculture
- qa
- ghana
- farmer-qa
- african-languages
- low-resource
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# Ghana Farmer Q&A

**151,469 question-answer pairs on Ghanaian smallholder farming, grounded in
15,174 recorded interviews with Ghanaian farmers.**

Every Q&A pair traces back to something a farmer actually said. The pairs are not
free-floating model output about African agriculture -- each one was written
against a specific interview passage, and that passage travels with it in the
`source_text` and `original_twi` columns, so you can always check an answer
against its source.

## Where the content comes from

1. **Field recordings.** Ghanaian farmers were recorded discussing their practice
   in Twi -- cocoa spacing, pest pressure, soil fertility, harvest and pricing.
   These were transcribed and translated into English, published as
   [`ghananlpcommunity/twi-english-agric`](https://huggingface.co/datasets/ghananlpcommunity/twi-english-agric).
2. **Question generation.** For each of the 15,174 interview passages, a model
   (Gemini 3.6 Flash) was asked to write ~10 standalone questions a farmer might
   put to an extension officer about the practices described in *that* passage,
   plus an explanatory answer for each. Each question had to stand on its own --
   no follow-ups, no references to "the text above" -- so the pairs are usable
   outside their original context.
3. **Flattening.** One row per Q&A pair, carrying its source passage forward.

So the *knowledge* is farmer-sourced and the *question-answer framing* is
model-written. That distinction matters for how you use it -- see Limitations.

## Columns

| column | description |
|---|---|
| `question` | Farming question in the first person, as if a farmer asking an extension officer |
| `answer` | Explanatory answer grounded in the source passage |
| `category` | Topic label chosen by the model, e.g. "Pest & Disease Control" |
| `source_text` | The interview passage this pair was generated from, in English |
| `original_twi` | The original Twi transcription of that passage |
| `file_name` | Source audio file identifier |

Questions average 16 words, answers 34 words. Common topics include pest and
disease control, soil and nutrient management, planting and spacing, crop growth
and behaviour, root and tuber crops, and vegetable farming.

## What it's useful for

- **Training and evaluating agricultural QA and extension-advice assistants** for
  Ghanaian farmers, in a register that matches how farmers actually ask.
- **Instruction tuning** where you want grounded, domain-specific QA rather than
  generic web text.
- **Retrieval and grounding research** -- `source_text` gives every answer a
  provenance passage, so you can measure whether an answer is supported.
- **A seed corpus for Ghanaian-language work.** Parallel Ewe, Ga and Twi versions
  are published separately (below).
- **Human post-editing** into a reviewed, gold-standard extension corpus. This is
  the highest-value next step and the dataset is structured for it.

## Parallel Ghanaian-language versions

- [`ghananlpcommunity/ghana-farmer-qa-ewe`](https://huggingface.co/datasets/ghananlpcommunity/ghana-farmer-qa-ewe) -- Ewe
- [`ghananlpcommunity/ghana-farmer-qa-ga`](https://huggingface.co/datasets/ghananlpcommunity/ghana-farmer-qa-ga) -- Ga
- [`ghananlpcommunity/ghana-farmer-qa-twi`](https://huggingface.co/datasets/ghananlpcommunity/ghana-farmer-qa-twi) -- Twi

## Limitations

- **The answers are model-written, not expert-reviewed.** They are grounded in
  farmer interviews, but no agronomist verified them. Do not deploy this as
  extension advice without review.
- **`category` is free text, not a controlled vocabulary.** The model chose a
  label per batch, producing ~24,000 distinct strings with heavy near-duplication
  ("Pest & Disease Control", "Pest and Disease Management", "Pest Management").
  Cluster or remap it before using it as a label.
- **Interview coverage shapes topic coverage.** Crops and regions that came up
  often in the recordings are well represented; others barely appear. This is not
  a balanced survey of Ghanaian agriculture.
- **The English passages are themselves translations** from Twi speech, so
  transcription and translation error is upstream of everything here.

## License

CC BY-NC 4.0, inherited from the source dataset
[`ghananlpcommunity/twi-english-agric`](https://huggingface.co/datasets/ghananlpcommunity/twi-english-agric).