CoolFace
Datasetpublic

kuzaai/kuza_sft_adversarial

Kuza SFT Adversarial Supervised fine-tuning data for Kuza, an offline agricultural assistant for smallholder farmers and agricultural extension workers in East Africa (English and Swahili). This repository is one of four Kuza SFT datasets. Dataset description This is a new hand-authored safety set. It is not derived from FarmerChat and is not one of the previously published Kuza agricultural Q&A corpora (kuzaai/agri_sft_prod_56k, kuzaai/agri_sft_prod_dedup_25k… See the full description on the dataset page: https://huggingface.co/datasets/kuzaai/kuza_sft_adversarial.

sourceHugging Facemitupdated 7d agoView on Hugging Face
0likes68downloads
Dataset Card

Kuza SFT Adversarial

Supervised fine-tuning data for Kuza, an offline agricultural assistant for smallholder farmers and agricultural extension workers in East Africa (English and Swahili). This repository is one of four Kuza SFT datasets.

Dataset description

This is a new hand-authored safety set. It is not derived from FarmerChat and is not one of the previously published Kuza agricultural Q&A corpora (`kuzaai/agri_sft_prod_56k`, `kuzaai/agri_sft_prod_dedup_25k`, `kuzaai/agri_sft_25k_swahili`).

Prompts try to elicit unlabeled pesticide doses, jailbreaks, leftover-chemical recipes, and illegal agrochemical shopping. Assistant turns refuse, decline invented rates, and redirect to a locally registered product label or an extension officer. Do not treat adversarial prompts as permitted behavior; only the refusal responses are the training target.

Kuza supervised fine-tuning samples 5% of the English train count from this split.

  • —Rows: 130
  • —Split: train only
  • —License: MIT

Safety

User turns are intentionally harmful-looking. They are included so a farm assistant learns to refuse unsafe agrochemical requests, not to teach those behaviors. Do not treat prompts as advice. Responses are refusals, not mixing instructions.

Languages

languagerows
english65
swahili65

Dataset structure

All four Kuza SFT repos share this schema:

fieldtypedescription
messageslist of {role, content}Full dialog. Roles are user and assistant. English and Swahili rows are typically 2 turns; multi-turn rows are 4 turns.
instructionstringLast user turn (same text SFT uses as the prompt side of the pair).
responsestringLast assistant turn.
languagestringenglish or swahili.
sourcestringOrigin bucket (english, swahili, adversarial, or multiturn).
source_idstringStable row id from prep, or {source}-{index} when the local file had none.

Splits

splitrows
train130

How to load

python
from datasets import load_dataset

ds = load_dataset("kuzaai/kuza_sft_adversarial", split="train")
print(ds[0]["messages"])

Intended use

Fine-tune a chat model as an East African farm assistant: direct, specific answers in the user's language, with concrete steps rather than generic advice, and without inventing pesticide or veterinary doses.

This split is meant to be mixed with the sister Kuza datasets below, not used as a complete training mixture by itself.

Limitations

  • —Advice is not a substitute for a local product label, veterinarian, or extension officer. Rates, products, and legal rules vary by country.
  • —Domain is English and Swahili agriculture in East Africa. It is not a general instruction corpus.

License

MIT License. You may use, copy, modify, merge, publish, distribute, and sublicense this dataset.

Sister datasets