DYNA-SKILL

Dynamic Self-Prompting Knowledge Graphs for Improving Logical Reasoning in Language Models

Jongwon Ryu, Mingi Kim, and Junyeong Kim

Research code for constructing and processing the self-prompted knowledge graphs introduced in the paper.

Overview

DYNA-SKILL links predefined commonsense relations with context-dependent relations generated through self-prompting. Each record connects two triples:

Head -- Predefined Relation --> Tail -- Dynamic Relation --> Additional Tail

The paper combines 35 predefined relations with 133 dynamic relations. This repository releases the graph-generation prompts and data-processing scripts; the complete generated corpus, fine-tuning/evaluation code, and model checkpoints are not included.

Setup

Use Python 3.10 or 3.11 and an isolated environment:

git clone https://github.com/jongwonryu/dyna-skill.git
cd dyna-skill
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

The research scripts use openai==0.28.1, which supports their original ChatCompletion interface. Configure credentials through the environment, not by editing source files:

export OPENAI_API_KEY="YOUR_API_KEY"
export OPENAI_MODEL="gpt-4-turbo"
python scripts/check_api.py

gpt-4-turbo is the historical model default, not a guarantee of current account availability. Set OPENAI_MODEL to an accessible Chat Completions model with compatible parameters. The API check and generation scripts make paid requests; changing models may change the generated data and does not reproduce the original paper's results.

Pipeline

Run commands from the repository root. The generation prompts and head categories are defined in scripts/generate_graph.py.

# Generate nested dual-triple records. Review the scope before starting:
# the full configuration makes many paid API requests.
python scripts/generate_graph.py

# Split each record into predefined and dynamic triples, preserving pair IDs.
python scripts/flatten_graph.py --input example.json --output example_triples.json

# Apply the original relation-frequency filtering and sentence templates.
python scripts/clean_relations.py
python scripts/triples_to_text.py

Stage	Input	Output
Graph generation	Head categories and relation prompts	example.json
Flattening	example.json	example_triples.json
Relation cleaning	example_triples.json	cleaned_triples.json
Text conversion	cleaned_triples.json	converted_text_data.txt

scripts/extend_tails.py is an optional legacy expansion stage. It reads example.json and writes example_updated.json; pass that output to the flattening script if using the expanded records. The original cleaning rule retains relations whose raw label occurs more than 10 times, then removes punctuation.

Generation scripts write progress files in the working directory. They are research utilities, not a transactional job system: retain backups of outputs and state before rerunning an interrupted generation job.

Data Format

The generated JSON groups records by category and predefined relation:

{
  "Social-Interaction Relations": {
    "xIntent": [
      {
        "Head": "A person studies for an exam",
        "Tail": "To understand the course material",
        "Dynamic Relation": "Supports",
        "Additional Tail": "Preparing a study plan"
      }
    ]
  }
}

This schema example explains the format; it is not a released experimental sample. Flattened records use Head, Relation, and Tail, plus Category, Pair ID, and Triple Type so the two linked triples can be traced back to their source.

Citation

@article{ryu2025dynaskill,
  title   = {{DYNA-SKILL}: Dynamic Self-Prompting Knowledge Graphs for Improving Logical Reasoning in Language Models},
  author  = {Ryu, Jongwon and Kim, Mingi and Kim, Junyeong},
  journal = {IEEE Access},
  volume  = {13},
  pages   = {188326--188334},
  year    = {2025},
  doi     = {10.1109/ACCESS.2025.3626479}
}

Contact

Jongwon Ryu: fbwhddnjs511@cau.ac.kr.

No software license has been specified for this repository; the paper's publication license does not automatically license the code.