ogma-prompt-injection-benign

~11,400 benign prompts (label 0) for training and evaluating prompt-injection classifiers. They are realistic, everyday requests a user would send an AI assistant — the kind a naive classifier wrongly flags as attacks.

Built to fix a specific failure: classifiers trained on prompt-injection datasets often learn that any instruction looks malicious, because the benign side of those datasets is templated filler. This set provides genuine, varied benign instructions instead.

Coverage

Synthetically generated with Claude Haiku, spanning:

• many task types (summarise, translate, rewrite, draft, explain, compare, classify, brainstorm, …) and everyday domains,
• both imperatives and questions,
• a range of lengths, from one-liners to long requests with pasted context,
• abstract / wordplay micro-tasks (name/generate/construct, fill-in-the-blank, quizzes, portmanteaus) — a category general benign data tends to miss,
• shell commands and code (subtype command / code) — ordinary developer commands and snippets, persona-driven, from rich pipelines (git status, npm install, docker ps -a, find … -name) down to trivial bare utilities (pwd, whoami, cat file.txt, echo 'Hello'). Added because classifiers otherwise learn that command syntax itself means attack, breaking tool/agent guardrails that screen commands before running them.

Schema

field	meaning
text	the benign prompt
label	always 0 (benign)
form	question or statement
length	short / medium / long_with_context
task_type, domain	generation metadata
subtype	natural / abstract / command / code

Usage

from datasets import load_dataset
benign = load_dataset("axiotic/ogma-prompt-injection-benign", split="train")

Pair with real injection examples to train a classifier. Trained model: axiotic/ogma-prompt-injection.

All rows are benign by construction; a light filter drops any that resemble injections, so the label is reliable.