Model Card for CDT-Cognition-Tagger

This model is a key component of the Cognition-Domain-Task (CDT) framework, a comprehensive capability framework for Large Language Models presented in our paper CDT: A Comprehensive Capability Framework for Large Language Models Across Cognition, Domain, and Task. It is specifically fine-tuned to classify a given instruction into one of 18 cognitive abilities defined by the CDT framework.

Model Details

Model Description

The Cognition dimension of the CDT framework is inspired by the Cattell-Horn-Carroll (CHC) theory of cognitive abilities, adapted for the context of LLMs. This model analyzes an instruction and identifies the primary cognitive skills required to fulfill it.

• Model type: Qwen2ForCausalLM
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model: Qwen2.5-7B-Base

Model Sources

• Repository: https://github.com/Alessa-mo/CDT
• Paper Link: https://arxiv.org/abs/2509.24422

Basic Usage

Please refer to https://github.com/Alessa-mo/CDT. You can run the following scripts to tag the cognition labels.

cd tag_annotate
export CUDA_VISIBLE_DEVICES=0
python annotate.py \
    --data_path path/to/your/data \
    --output_dir path/to/output/dir \
    --model_path CDT-Cognition-Tagger \
    --prompt_file ./prompt/annotation_prompt.jsonl \
    --cognition_skill_file ./prompt/cognition.json \
    --domain_skill_file ./prompt/domain.json \
    --task_skill_file ./prompt/task.json \
    --tag_type "cognition" \
    --batch_size 32

Note: Make sure your data is a JSON file and has the following format:

[
    {
        "messages": [
            {
                "role": "user",
                "content": "xxxx"
            },
            {
                "role": "assistant",
                "content": "xxxx"
            }
        ]
    },
]

Citation

If you find this model useful, please cite:

@misc{mo2025cdtcomprehensivecapabilityframework,
      title={CDT: A Comprehensive Capability Framework for Large Language Models Across Cognition, Domain, and Task}, 
      author={Haosi Mo and Xinyu Ma and Xuebo Liu and Derek F. Wong and Yu Li and Jie Liu and Min Zhang},
      year={2025},
      eprint={2509.24422},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.24422}, 
}