lalalamdbf/AgentSafeEvo-Attacker-SFT-Data

AgentSafeEvo Attacker SFT Data

This dataset contains 20,322 supervised fine-tuning examples for initializing the attacker in AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety. It trains the attacker to construct benign and adversarial agent tasks from an agent scenario, an intent, and available tool definitions.

Data Construction

The dataset is built from selected TOUCAN tasks. Gemini 3 Flash Preview is used to synthesize adversarial tasks according to specified attack intents. The resulting training data combines benign and red-team examples.

Example type	Count
Benign	10,012
Red-team	10,310
Total	20,322

Data Format

redteam_sft.json contains a JSON array in Alpaca format. Each example has four string fields:

Field	Description
system	Instructions for constructing the agent task.
instruction	The agent scenario, benign or harmful intent, and available tool definitions.
input	An empty string in this release.
output	The target task construction, including a construction trace, a user query, and tool functions with mocked calls and simulated replies.

The dataset can be used with LLaMA-Factory by registering these fields in dataset_info.json, including the system column mapping.

Loading

from datasets import load_dataset

dataset = load_dataset(
    "lalalamdbf/AgentSafeEvo-Attacker-SFT-Data",
    data_files={"train": "redteam_sft.json"},
    split="train",
)

Intended Use and Limitations

The dataset supports attacker initialization and research on red-teaming and LLM agent safety. It includes adversarial instructions and simulated tool content. Mocked tool calls and replies may differ from the behavior of real tools and deployments.