flamiinngo/agronomy-qa-agriculture

Agronomy QA Pairs — Agriculture Instruction Dataset

A concise, English-language question-and-answer dataset covering practical agriculture: crop management and planting, soil health and fertility, irrigation, and pest and disease control, with a smaller amount of livestock content. Curated for instruction fine-tuning of language models in the agriculture domain.

Rows	1,914 question/answer pairs
Unique questions	1,914 — no duplicate questions
Unique answers	1,870 (44 answers recur across distinct questions)
Nulls / empty fields	none
Question length	median 10 words (range 2–41)
Answer length	median 27 words, mean 34 (range 5–270)
Format	CSV — question, answer
Language	English
License	Apache 2.0

Answer length distribution

Under	Share of rows
20 words	34%
40 words	68%
60 words	89%
100 words	98%

Median answer is 180 characters. The long tail reaches 270 words — a handful of multi-step procedural answers (grain storage, disease treatment protocols) run long.

Topic coverage

Keyword-based, and overlapping — a single row can touch several topics.

Topic	Rows	Share
Crops & planting	995	52%
Pests & disease	685	36%
Soil & fertility	559	29%
Irrigation & water	470	25%
Livestock	74	4%

Livestock is thin — 4% of rows. Treat this as a crop-and-soil dataset with incidental livestock content, not a balanced agriculture corpus.

Why the answers are short

This is the defining property of the dataset. Answers are brief on purpose — a median of 27 words. Training a model on concise, real farming answers rather than long generated essays is what took the companion model from losing to its baseline to beating it 82 to 19 on a held-out agriculture benchmark.

If you are fine-tuning for a factual-evaluation setting, that brevity is a feature. If you need long-form explanatory output, this is the wrong dataset.

Usage

from datasets import load_dataset

ds = load_dataset("flamiinngo/agronomy-qa-agriculture", split="train")
print(ds[0])
# {'question': 'why is crop rotation important in farming?',
#  'answer': 'This helps to prevent soil erosion and depletion, and can also
#             help to control pests and diseases'}

Convert to chat format for instruction fine-tuning:

SYSTEM = "You are an agricultural advisor. Answer concisely and factually."

def to_chat(row):
    return {"messages": [
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": row["question"]},
        {"role": "assistant", "content": row["answer"]},
    ]}

chat = ds.map(to_chat, remove_columns=ds.column_names)

Fields

Column	Type	Description
question	string	A practical agriculture question, phrased as a farmer or extension user would ask it.
answer	string	A concise, factual answer covering crops, soil, water, pests, or livestock.

Source & Attribution

Derived from KisanVaani/agriculture-qa-english-only (Apache 2.0), which curated real agricultural Q&A from farming forums, extension resources, and FAQ sources. This version filters to substantive answers, cleans the text, and formats it for instruction fine-tuning.

Trained model

flamiinngo/agronomy-llama-3.3-70b — Llama-3.3-70B-Instruct + LoRA, trained on this data. Beats its base model 82 to 19 on the held-out Agriculture evaluation (reproduced at 71 vs 29).

Also mirrored on Kaggle: dataset · weights

Intended Use

Fine-tuning and evaluating question-answering models for agricultural guidance. Created as part of the Adaption Labs AutoScientist Challenge (Agriculture category).

Limitations

• Regional skew. The upstream corpus draws largely on Indian farming forums and extension resources. Crop varieties, growing seasons, and pest pressures reflect that context and may not transfer directly elsewhere. Some answers reference specific local seed varieties or district-level programmes by name.
• Topic imbalance. Crops and planting dominate at 52% of rows; livestock is only 4%.
• Answers are unverified at the item level. They come from real extension and forum sources and were filtered for substance, not fact-checked individually by an agronomist.
• Brevity limits nuance. Some questions genuinely warrant caveats the short answers omit.
• Informal register. Questions preserve the phrasing of real users, so capitalisation and grammar are inconsistent ("what is crop rotation", no question mark). This is intentional — it matches how the model will be queried — but it means the text is not editorially clean.
• 54 rows contain non-ASCII characters (curly quotes, degree signs, and similar). Read the CSV as UTF-8.
• English only.

Disclaimer

Content is for general informational and research purposes. Agricultural practices vary by region, crop, and climate; always consult local agricultural extension services for specific guidance.