README.md

metadata

license: cc-by-4.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: train_dataset.jsonl
      - split: validation
        path: eval_dataset.jsonl
      - split: test
        path: test_dataset.jsonl
  - config_name: full
    data_files:
      - split: full
        path: compiled_dataset.jsonl

SysMLv2 Repair with SLMs

Dataset used in "Automated Semantic Fault Localization in SysML v2: A Human-in-the-Loop Framework Using Knowledge-Graph Augmented LLMs", presented at INCOSE International Symposium 2026.

Dataset Structure

This dataset provides two configurations:

• default: Contains train/validation/test splits used for fine-tuning small models. Samples exceeding 2048 tokens have been removed.
• full: Contains complete dataset

Task

Given SysML v2 code and relevant context (compiler error, or relevant domain rules), identify and fix potential faults in it. The output is either the corrected code or a diff patch for correcting.

Fields

• id: Unique identifier for each dataset instance.
• source_id: Identifier of the original (clean) code example from which this instance was derived.
• mutation_category: High-level category of the applied mutation:
  • domain: domain-specific semantic changes
  • syntax: syntactic errors
  • none: no mutation applied
• mutation_type: Specific mutation operator used to generate the erroneous code.
• bad_code: SysML v2 code potentially containing injected errors.
• good_code: Correct version of the code.
• diff_patch: Unified diff representing the transformation from bad_code to good_code.
• base_prompt: Base prompt template used for fine-tuning; contains bad_code and compiler error in case of syntax mutations.
• prompt: Prompt containing additional context (e.g., relevant domain rules), for domain and none mutations, where no compiler error occurs.
• code_response: Repair in full code form.
• patch_response: Repair in diff/patch format.
• length: Total number of tokens in the full training sequence (prompt + response).

Data Creation

1. A seed set of 256 examples was created from a combination of public and author-generated SysML v2 code.
2. Synthetic errors were introduced via:
   • Syntactic mutations (5,497 instances)
   • Domain/semantic mutations based on violations of rules defined in a knowledge graph (1,402 instances)
3. An equal number of correct (unmutated) examples were included to support classification of correct vs. erroneous code, resulting in 8,301 instances.
4. Additional context was generated for each example:
   • Compiler error messages for syntactic errors
   • Relevant domain rules for semantic cases
5. Target outputs were derived, including corrected code and corresponding diff patches.
6. Entries longer than 2048 tokens were dropped, and remaining (7,780 instances) were split into train/validation/test sets using a 70/15/15 ratio.

License

This dataset is released under CC BY 4.0. Attribution to the original authors is required.

Citation

GitHub Repository: SysMLv2 Repair with KG-SLMs

@inproceedings{alshami2026sysml,
  title={Automated Semantic Fault Localization in SysML v2: A Human-in-the-Loop Framework Using Knowledge-Graph Augmented LLMs},
  author={Al-Shami, Haitham and Malik, Rohail and Ala-Laurinaho, Riku and Veps{\"a}l{\"a}inen, Jari and Viitala, Raine},
  booktitle={Proceedings of the 36th INCOSE International Symposium},
  year={2026},
  address={Yokohama, Japan},
  month={June},
  date={16}
}