vicharanashala/ajrasakha-dataset-v1

Ajrasakha Dataset v1

A question–answer dataset of Indian agriculture advisory queries and expert-reviewed answers, primarily in English with occasional local-language terms and English/local-language code-mixing (e.g. Hindi, Marathi, Telugu, Tamil, Punjabi, Bengali script borrowed words). Built from the production data of Ajrasakha, an AI-assisted agricultural advisory platform by annam.ai.

At a glance

Source	annam.ai — Ajrasakha platform
Coverage	Up to 06 September 2026
Total Records	40,065
Languages	Primarily English; with local-language terms and English/local-language code-mixing (Hindi, Marathi, Telugu, Tamil, Punjabi, Bengali, etc.)
Domain	Agriculture (crop protection, nutrients, agronomy, varieties, market info, weather, horticulture)
License	MIT
Splits	train (39,265) / validation (400) / test (400)

Schema

Each row is one question. The question and its best (final / approved) answer are denormalised into a single record.

Column	Type	Coverage	Description
question_id	string	100%	Ajrasakha question identifier
question	string	100%	The question text
priority	string	100%	high, critical, or medium
source	string	100%	Ingestion channel: AGRI_EXPERT, AJRASAKHA, MANUAL, WHATSAPP, OUTREACH
state	string	100%	Indian state declared by the farmer
district	string	100%	District declared by the farmer
crop	string	100%	Crop as declared by the farmer (raw, may have casing variants)
normalised_crop	string	100%	Crop normalised to a canonical value
season	string	100%	Kharif, Rabi, Zaid, Perennial, General, etc.
domain	string	100%	Advisory domain: Plant Protection, Pest, Disease, Fertilizer and Nutrient, Agronomy, Variety, Horticulture, Market Information, Weather, etc.
user_id	string	73%	Opaque identifier of the farmer who asked
question_created_at	string	100%	ISO 8601 timestamp
final_answer_id	string	92%	Identifier of the chosen answer
final_answer	string	92%	The answer text
final_answer_approval_count	float	92%	Number of approvals the answer received
final_answer_author_id	string	90%	Opaque identifier of the expert who drafted the answer
final_answer_approved_by	string	82%	Opaque identifier of the reviewer who approved it
final_answer_sources	string	92%	JSON array of citation objects: {source, sourceType, page, sourceName}

Rows without a final answer were either closed as duplicates, not approved, or non-agricultural.

Answer selection

For each unique question, the best available answer is chosen in this order:

1. isFinalAnswer = true AND status = approved
2. Any status = approved answer (highest approvalCount first)
3. Highest approvalCount regardless of status

The non-embedded fields final_answer_sources and final_answer are the model's outputs from the underlying pipeline; internal-only fields (moderator routing, review timelines, embeddings, internal file paths) are not included in this release.

Sources / citations

Each answer is grounded in agricultural reference material — primarily state Packages of Practices (POP), ICAR / NIPHM / IPM guidelines, and government scheme documents. Citations are preserved as JSON in final_answer_sources. Each entry has:

• source — public URL to the document
• sourceType — central (Indian government) or state
• page — page number in the source PDF
• sourceName — human-readable document title

Use cases

• Fine-tuning agricultural question-answering models for Indian English with local-language code-mixing
• Retrieval-augmented generation (RAG) over Package-of-Practices style documents
• Evaluation of factual grounding in low-resource, multilingual agricultural domains
• Studying farmer query patterns and crop-distribution across Indian states

Limitations

• Coding and priority are platform-specific and may not generalise
• Question and answer quality varies with the underlying expert process
• Some crops are repeated with inconsistent casing in the raw crop field; use normalised_crop for statistics
• Identifier columns (user_id, final_answer_author_id, final_answer_approved_by) are opaque — they cannot be resolved to public profiles and are not suitable for individual-level analysis
• The dataset reflects the production distribution of farmer queries and is not a balanced sample

Citation

If you use this dataset in research or product work, please cite:

Ajrasakha Dataset v1. annam.ai. 2026. https://huggingface.co/datasets/vicharanashala/ajrasakha-dataset-v1

License

Released under the MIT License. You are free to use, modify, and distribute, including for commercial purposes, subject to the standard MIT terms.

About

Maintained by annam.ai — an AI platform for Indian agriculture that connects farmers with verified, source-grounded agricultural advice.