cdreetz/on-call-agent-grpo-dataset

On-Call Agent GRPO Dataset

This dataset contains synthetic incident scenarios for training on-call engineering agents using Group Relative Policy Optimization (GRPO). Each example represents a different incident type with corresponding system state data that agents can investigate using diagnostic tools.

Dataset Description

The dataset simulates realistic incident scenarios that on-call engineers encounter in production environments. Each incident includes:

• Incident alerts with varying severity and descriptions
• System state data including status pages, Slack messages, deployment history, and application logs
• Root cause diagnoses for supervised learning and reward computation

Dataset Structure

Each example contains the following fields:

• question (string): The incident alert description presented to the agent
• answer (string): The expected root cause diagnosis
• state (dict): Complete system state data including:
  • status_pages: Service health status from monitoring dashboards
  • slack: Communication logs from team channels
  • deployments: Recent deployment history with success/failure status
  • logs: Application logs with timestamps and error messages
• issue_type (string): Category of incident (status_page, slack, deployment, logs)

Incident Types

The dataset includes four main categories of incidents:

1. Status Page Issues (50%): Service degradations visible on monitoring dashboards
2. Slack Communicated Issues (20%): Problems reported through Slack/communication channels
3. Deployment Failures (20%): Issues introduced by recent code deployments
4. Application Errors (10%): Runtime errors visible in application logs

Example Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("cdreetz/on-call-agent-grpo-dataset")

# Example incident
example = dataset['train'][0]
print("Incident:", example['question'])
print("Expected diagnosis:", example['answer'])
print("Available system data:", example['state'].keys())

Training Agents

This dataset is designed for training diagnostic agents that can:

1. Analyze incident alerts and determine investigation priorities
2. Use diagnostic tools to query system state (status pages, logs, etc.)
3. Synthesize information from multiple sources to identify root causes
4. Provide actionable diagnoses for incident resolution

Recommended Tools

Agents should be equipped with these diagnostic capabilities:

• check_status_pages(service): Query service health status
• search_slack_messages(query): Search team communications
• check_deployment_failures(hours_back): Review recent deployments
• query_logs(error_pattern): Search application logs

Training Framework

This dataset is compatible with the Verifiers framework for GRPO training:

import verifiers as vf
from datasets import load_dataset

# Load dataset
dataset = load_dataset("cdreetz/on-call-agent-grpo-dataset")

# Create environment with diagnostic tools
env = IncidentAnalysisEnv(
    dataset=dataset,
    tools=[check_status_pages, search_slack_messages, 
           check_deployment_failures, query_logs]
)

# Train with GRPO
trainer = vf.GRPOTrainer(model=model, env=env, args=training_args)
trainer.train()

Data Generation

The dataset was generated using a systematic approach:

1. Base state generation: Random but realistic system states with healthy services
2. Issue injection: Targeted problems introduced into specific data sources
3. Narrative creation: Incident descriptions matching the injected issues
4. Randomization: Varied timing, services, and error patterns for diversity

All incidents include items in each data category (status pages, Slack messages, deployments, logs) with only one containing the relevant diagnostic information.

Dataset Statistics

• Total examples: 1,000 incident scenarios
• Average question length: ~85 characters
• Average answer length: ~65 characters
• Issue distribution:
  • Status page: 50%
  • Slack/communications: 20%
  • Deployment: 20%
  • Application logs: 10%

Evaluation

Agent performance can be evaluated on:

• Diagnostic accuracy: Correct identification of root causes
• Tool efficiency: Minimal tool calls while gathering necessary information
• Investigation methodology: Systematic approach to incident analysis

Limitations

• Synthetic data: Generated scenarios may not capture all real-world complexity
• Simplified state: System state is limited to four data sources
• English only: All incident descriptions and logs are in English
• Deterministic issues: Each incident has exactly one clear root cause

Citation

If you use this dataset in your research, please cite:

@dataset{on_call_agent_grpo_2025,
  title={On-Call Agent GRPO Dataset},
  author={cdreetz},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/cdreetz/on-call-agent-grpo-dataset}
}

License

This dataset is released under the MIT License. See LICENSE for details.

Contributing

To report issues or suggest improvements to the dataset, please open an issue in the repository.