faodl/amis-agri-wheat

AMIS Wheat Commodity Media Dataset

Dataset Summary

This dataset contains text chunks extracted from multilingual news articles and annotated for wheat commodity relevance.

The annotations were generated using GPT-5 nano as an automated labeling model, following a structured prompt-based classification workflow.

Each text chunk was evaluated to determine whether its main content is primarily related to the wheat commodity, including topics such as production, trade, prices, supply, demand, stocks, weather impacts, and market developments.

The resulting labels are intended to support the development and evaluation of binary text classification models for commodity relevance detection.

The objective is to support the development and evaluation of binary text classification models that identify whether a text excerpt is primarily about the wheat commodity, regardless of the specific thematic dimension.

This includes, but is not limited to:

• production
• trade
• exports and imports
• prices and markets
• stocks and inventories
• weather and crop outlook
• policy and regulation
• logistics and supply chain
• food security and market disruptions

The dataset was developed as part of the AMIS (Agricultural Market Information System) media analysis workflow at the FAO Data Lab.

Each record corresponds to a text chunk extracted from a news article.

The target variable is binary:

• 1 → chunk is about the wheat commodity
• 0 → chunk is not about wheat

This dataset is intended for:

• supervised text classification
• weak supervision validation
• commodity-specific media monitoring
• benchmarking multilingual NLP models

---

Dataset Structure

The dataset is split into train, validation, and test partitions.

Split	Rows
Train	3,622
Validation	759
Test	762
Total	5,143

Features

Feature	Type	Description
chunk_id	string / int	Unique identifier of the chunk
id	string / int	Source article identifier
chunk_text	string	Text chunk extracted from article
label	int	Binary target label (0/1)

---

Label Definition

Positive class (label = 1)

The chunk is primarily about the wheat commodity.

This includes text related to:

• wheat production
• harvest and planting
• wheat prices
• futures and commodity markets
• exports and imports
• global supply and demand
• storage and stocks
• crop forecasts
• drought and weather impacts
• policy affecting wheat markets

Negative class (label = 0)

The chunk is not primarily about wheat.

Examples include:

• other commodities (e.g. maize, rice, soybeans)
• general macroeconomic news
• politics unrelated to wheat
• unrelated agriculture sectors
• logistics with no wheat reference

---

Data Collection and Processing

The dataset was derived from multilingual news articles retrieved from the FAO Data Lab News Database.

The workflow includes:

1. retrieval using wheat-related seed keywords
2. keyword expansion and multilingual augmentation
3. chunk extraction and preprocessing
4. annotation and validation
5. grouped train / validation / test split

Where possible, article-level grouping was preserved during splitting to reduce data leakage across partitions.

---

Intended Uses

This dataset is intended for:

• commodity relevance classification
• multilingual text classification
• transformer fine-tuning
• active learning workflows
• weak supervision validation
• media monitoring systems

Example applications:

• AMIS commodity monitoring
• event detection pipelines
• market intelligence systems
• early warning analytics

---

Limitations

Some limitations should be considered:

• media coverage bias
• retrieval bias from seed keywords
• ambiguity in cross-commodity articles
• multilingual imbalance
• evolving market terminology

This dataset should not be interpreted as an official wheat market statistic.

---

Citation

If using this dataset, please cite the FAO Data Lab / AMIS media analysis workflow.

@dataset{amis_wheat_commodity_dataset,
  title={AMIS Wheat Commodity Media Dataset},
  author={FAO Data Lab},
  year={2026}
}

Contact

For methodology questions, please contact the FAO Data Lab.