Lelonthecodeur's picture
Publish ATRD 1M dataset: Parquet, README, MIT License and tags
98b3eb6 verified
|
Raw History Blame Contribute Delete
4.07 kB
metadata
license: mit
language:
  - en
pretty_name: Agent Training Reasoning Dataset Dense
task_categories:
  - text-generation
size_categories:
  - 1M<n<10M
tags:
  - agentic-ai
  - agent-training
  - reasoning
  - tool-use
  - synthetic-data
  - parquet
  - multi-step-planning
  - decision-making
  - data-analysis
  - code-generation
  - workflow-automation
  - problem-solving
  - verification
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet

Agent Training Reasoning Dataset Dense

A synthetic dataset for general-purpose agentic AI training.

Repository: Lelonthecodeur/agent-training-reasoning-dataset-dense

Overview

Agent Training Reasoning Dataset Dense (ATRD) is a programmatically generated dataset designed to provide structured examples across multiple agentic task domains.

It focuses on task understanding, planning, tool selection, verification, decision-making and structured responses.

Dataset statistics

Property Value
Total examples 1,000,000
Domains 20
Difficulty levels 4
Format Apache Parquet
Compression Zstandard
Parquet shards 20
Generator version 1.0.0
Random seed 42
Language English
Data type Synthetic

Domain coverage

  • planning
  • coding
  • research
  • data_analysis
  • mathematics
  • tool_selection
  • verification
  • project_management
  • information_retrieval
  • problem_solving
  • communication
  • safety
  • decision_making
  • debugging
  • scheduling
  • file_management
  • summarization
  • critical_thinking
  • workflow_automation
  • ambiguity_resolution

Difficulty distribution

  • advanced: 250000
  • basic: 250000
  • expert: 250000
  • intermediate: 250000

Dataset schema

Each example contains:

Field Description
id Unique example identifier
messages System, user and assistant messages
domain Task domain
difficulty Difficulty category
available_tools Tools available in the scenario
synthetic Whether the example is synthetic
generator_version Generator version

The assistant message contains a structured JSON object with a proposed plan, actions, simulated tool trace, final answer and verification metadata.

Loading the dataset

Install the required libraries:

pip install datasets pyarrow

Load the dataset:

from datasets import load_dataset

dataset = load_dataset(
    "Lelonthecodeur/agent-training-reasoning-dataset-dense",
    split="train"
)

print(dataset)
print(dataset[0])

Generation methodology

Examples are generated programmatically from domain-specific templates and randomized variables.

IDs are deterministic and unique within this generation run.

The dataset is divided into Parquet shards and compressed using Zstandard.

Important limitations

This is a synthetic baseline dataset.

  • Tool traces are simulated placeholders, not real tool executions.
  • Plans and responses are template-generated.
  • Examples have not been independently human-verified.
  • The dataset may contain repetitive patterns.
  • Difficulty labels are assigned by a deterministic cycle, not by independent difficulty evaluation.
  • Mathematical examples are limited to simple arithmetic.

Do not interpret the dataset as one million independently verified reasoning trajectories.

Independent quality evaluation, deduplication, diversity testing and benchmark validation are recommended before production training.

Intended uses

  • Agentic AI research
  • Tool-selection experiments
  • Planning and workflow experiments
  • Synthetic data pipeline development
  • Dataset engineering experiments

License

Released under the MIT License.

See the LICENSE file for the complete license text.

Citation

@misc{lelon_agent_training_reasoning_dataset,
  title={Agent Training Reasoning Dataset Dense},
  author={Lelonthecodeur},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/Lelonthecodeur/agent-training-reasoning-dataset-dense}
}

Maintainer

Lelonthecodeur

Dataset version: 1.0.0