Datasetpublic

Alabi-Ayobami/naija-agric-qa-synth

AgriPadi Synthetic SFT Production-ready synthetic agriculture-advisory SFT data for West Africa (English, Nigerian Pidgin, Hausa), generated by the AgriPadi pipelines. Pipeline source (reproducible): https://github.com/ThatLinuxGuyYouKnow/agripadi_synth_pipeline All rows are published in the regular chat format (messages as user/assistant pairs). The raw-completion surface used for lm-eval profiling is intentionally excluded. Composition 6,701 training rows:… See the full description on the dataset page: https://huggingface.co/datasets/Alabi-Ayobami/naija-agric-qa-synth.

sourceHugging Facemitupdated 4d agoView on Hugging Face

0likes25downloads

Dataset Card

AgriPadi Synthetic SFT

Production-ready synthetic agriculture-advisory SFT data for West Africa (English, Nigerian Pidgin, Hausa), generated by the AgriPadi pipelines.

Pipeline source (reproducible): https://github.com/ThatLinuxGuyYouKnow/agripadi_synth_pipeline

All rows are published in the regular chat format (messages as user/assistant pairs). The raw-completion surface used for lm-eval profiling is intentionally excluded.

Composition

6,701 training rows:

—advisory: 737
—mc_arc: 2221
—mc_letter: 2221
—short_answer: 1522

Languages

—English: 6381
—Hausa: 320

Subdomains

—crops_agronomy: 1105
—livestock_poultry_fish: 961
—markets_agribusiness: 478
—pest_disease_weeds: 1741
—policy_extension: 360
—postharvest_storage: 607
—soil_fertilizer_irrigation: 813
—weather_climate: 636

Schema

Field	Type	Notes
`id`	string	Unique row id (`source_id:variant`)
`source_id`	string	Stage-1 source row id
`variant`	string	`mc_letter`, `mc_arc`, `short_answer`, or `advisory`
`surface`	string	Always `chat` in this release
`subdomain`	string	Curriculum subdomain
`track`	string	`excerpt` (grounded) or `grid` (breadth)
`language`	string	`English` or `Hausa` (from the stage-1 source row)
`messages`	list	Chat turns: `{role, content}`
`question`	string	Advisory only: the farmer message
`expected_behavior`	string	Advisory only: `answer_directly`, `clarify`, or `reassure`
`goal`	string	Advisory only: farmer intent label
`verified`	bool	Advisory only: LLM verifier verdict
`verifier_note`	string	Advisory only: verifier comment, if any

MC rows appear as two variants per source item — mc_letter (letter continuation, MMLU-style) and mc_arc (choice-text continuation, ARC-style) — so both loglikelihood serializations are covered.

Provenance & quality gates

Generated from a curated seed corpus of sourced extension excerpts, then gated by:

1.deterministic length-bias rejection on multiple-choice items
2.banned-actives screen
3.cross-track deduplication
4.source-voice flagging (excluded from this release)
5.independent LLM verification, with rejection-reason feedback repair

See the pipeline repository for the runbook, gate design notes, and repair passes.

Loading

python

from datasets import load_dataset

ds = load_dataset("Alabi-Ayobami/naija-agric-qa-synth", split="train")

Intended use

Supervised fine-tuning of small agricultural advisory assistants for West African farming contexts. Not a factual authority — always validate agronomic recommendations against local extension services.