README.md

metadata

license: mit
language:
  - en
tags:
  - code
  - Text
  - agent
pretty_name: Manage tasks with agents
size_categories:
  - 1M<n<10M

Agent Task Management Dataset

5,000,000 compact synthetic examples for agent task execution.

Repository: Lelonthecodeur/agent-task-management-dataset

Purpose

This dataset is designed to train general-purpose agents to plan, coordinate, execute, verify and report complex tasks involving multiple tools and domains.

Domains

• email_writing
• document_creation
• app_development
• terminal_operations
• backend_engineering
• financial_analysis
• data_analysis
• project_management
• research
• workflow_automation
• software_testing
• ui_ux_design
• file_management
• decision_making
• database_management
• api_integration
• debugging
• cross_domain_execution

Scenario types

The dataset includes:

• Successful task completion.
• Failed attempts.
• Error detection and recovery.
• Impossible or unsupported tasks.
• Clarification and ambiguity handling.

Example structure

Each Parquet record contains:

• id
• task
• messages
• domains
• difficulty
• outcome
• available_tools
• constraints
• expected_outputs
• execution_plan
• tool_trace
• error
• clarification
• verification
• final_result
• scenario_family
• split
• synthetic
• generator_version

The messages field contains a compact JSON conversation with system, user and assistant messages.

Dataset splits

The dataset uses approximately:

• 80% training
• 10% validation
• 10% test

Splits are sampled independently. Scenario families are not strictly held out across splits, so this dataset should not be treated as a leakage-proof benchmark.

Synthetic-data disclosure

All examples are programmatically generated.

Tool traces are simulated representations. No real terminal commands, emails, financial transactions, API requests, database changes or deployments were performed.

The examples do not establish that a task was actually completed in an external environment.

The generator uses structured templates and randomized combinations. Consequently, some examples may be repetitive, simplified or less realistic than expert-written trajectories.

Difficulty labels are assigned by the generator and are not independently benchmarked.

This dataset is a synthetic training resource, not a substitute for verified real-world agent execution data.

Intended use

• Agent planning and task decomposition.
• Multi-domain workflow training.
• Tool-selection reasoning.
• Error handling and recovery.
• Clarification behavior.
• Output verification.
• Structured task reporting.

Limitations

Do not interpret simulated execution traces as real execution evidence. Verify model behavior independently before deploying systems trained on this dataset.

Generation details

• Examples: 5,000,000
• Format: Parquet
• Compression: Zstandard
• Generator version: 1.0.0
• License: MIT

License

MIT License. See the LICENSE file.