FuzzyLabs/guardrail-scenarios

Prompt-Manipulation Scenarios

It is built from the raw source dataset FuzzyLabs/guardrail-scenarios-raw, itself transformed from reshabhs/SPML_Chatbot_Prompt_Injection (MIT).

This repository reshapes its rows into chat conversations with fixed train/val/test splits, ready for supervised finetuning, inference, and evaluation. No new labels are added.

	Raw source (FuzzyLabs/guardrail-scenarios-raw)	This dataset
Purpose	Archival source of truth	Model-ready training and evaluation data
Row format	Flat source/context/document/decision columns	id + messages
Prompt	None	System prompt and user template applied
Target	One three-way decision column	One JSON string in the assistant turn
Changes when	The source dataset is revised	The prompt, cleaning, or split settings change

Pin a specific revision of this repository, not main, so that finetuning, inference, and evaluation all use identical splits and prompts.

Files

• data/: the train, val, and test splits used for finetuning and inference
• raw.jsonl: every cleaned record with its turn, context, reference decision, and split, used as evaluation ground truth
• preparation_config.yaml: the exact cleaning, splitting, and prompt configuration used

Processing

• Source splits: train
• Turn length: 1 to 4,000 characters
• Deduplicated by turn identity (context and document) and split deterministically with seed 42
• 16,012 raw rows; 16,012 records survived cleaning

Splits

• train: 13,512 records
• val: 1,000 records
• test: 1,500 records

Schema

Each record contains a stable id and messages with system, user, and assistant turns. The assistant turn contains the reference decision as JSON (see GuardrailLabel).

Cite

@misc{sharma2024spml,
    title={SPML: A DSL for Defending Language Models Against Prompt Attacks},
    author={Reshabh K Sharma and Vinayak Gupta and Dan Grossman},
    year={2024},
    eprint={2402.11755},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}