Dataset Card for TMC-Llama-3.1-8B-Instruct Dataset

This is the official dataset source of the paper: DynamicMind: A Tri-Mode Thinking System for Large Language Models.

Code: The official GitHub repository.

Mind Router Models:

• MR-Llama-3.1-8B-Instruct
• MR-Qwen-2.5-7B-Instruct

Dataset Summary

The Thinking Mode Capacity (TMC) Dataset is designed to train a Mind Router, a lightweight classification model that predicts the optimal thinking mode for an LLM to use when answering a given question.

The dataset was created to support the DynamicMind framework, which enables an LLM to dynamically switch between three Thinking Modes:

• Fast: Delivers rapid, intuitive responses by limiting cognitive depth, prioritizing speed over thorough reasoning.  

• Normal: Leverages the LLM's native capabilities to balance response quality and efficiency.

• Slow: Executes deep analytical reasoning for high-quality outputs with increased computational costs.

The TMC dataset maps questions to the optimal thinking mode of an LLM. This optimal mode is determined by the Thinking Density metric, which calculates the best trade-off between response accuracy and token consumption for each question. By training on this dataset, a Mind Router learns to select the most efficient and effective reasoning strategy for the paired LLM.

Languages

The text in the dataset is in English.

Dataset Structure

Data Instances

{
    "input": "Please answer the following question with true or false, question: does the etihad stadium manchester have a roof?\n\nAnswer format: true/false",
    "task": "boolq",
    "output": "slow"
},

Data Fields

• input: string - The input question for the LLM.
• task: string - The source dataset of the input.
• output: string - The optimal thinking mode for the given question, determined by the Thinking Density metric. The possible values are fast, normal, or slow.

Dataset Creation

For the mathematical capability evaluation dataset, we extracted all data from the train splits of the GSM8K and MATH datasets. For the commonsense capability evaluation, we sampled 5,000 entries from the training split of eight benchmark datasets: BoolQ, PIQA, SIQA, HellaSwag, WinoGrande, ARC-e, ARC-c, and OpenBookQA. If a dataset had fewer than 5,000 entries, the entire train split was selected.

Licensing Information

The TMC dataset is licensed under the Apache-2.0 License.

Citation

If you find our work helpful, please cite our paper:

@article{li2025dynamicmind,
  title={DynamicMind: A Tri-Mode Thinking System for Large Language Models},
  author={Li, Wei and Wei, Yanbin and Huang, Qiushi and Yan, Jiangyue and Chen, Yang and Kwok, James T and Zhang, Yu},
  journal={arXiv preprint arXiv:2506.05936},
  year={2025}
}