Mainframebench Eval

mainframebench-eval

XMainframe: A Large Language Model for Mainframe Modernization — Dau et al. (2024) (arXiv:2408.04660, 2024)

What this evaluates

Probes large language models' ability to reason about legacy mainframe systems, interpret COBOL code, and generate accurate technical summaries. It tests domain-specific code understanding through multiple-choice questions, open-ended QA, and text generation tasks.

Datasets

• MainframeBench — total ?; splits: test (-1); repo https://github.com/FSoft-AI4Code/XMainframe

Metrics

• Accuracy (primary) — range: percent
  • Percentage of correctly selected options in multiple-choice questions.
• MAP — range: [0, 1]
  • Mean Average Precision for ranking or retrieving relevant answers in question answering.
• F1-Score — range: [0, 1]
  • Harmonic mean of precision and recall for generated answers.
• BERTScore — range: [0, 1]
  • Cosine similarity between contextual embeddings of generated and reference text using BERT.
• RougeL — range: [0, 1]
  • Recall of the longest common subsequence between generated and reference text.
• Meteor — range: [0, 1]
  • Metric for evaluation with alignment and synonymy, weighting recall and penalty for fluency.
• BLEU-4 — range: [0, 100]
  • 4-gram n-gram precision with brevity penalty comparing generated text to references.

Input / output format

Input: Multiple-choice questions with four options (A, B, C, D); open-ended questions about mainframe/COBOL systems; COBOL code snippets requiring natural language summaries.

Output: Single letter option (A/B/C/D) for MCQ; free-text answer for QA; concise natural language summary for COBOL code.

Scoring recipe

def score_mcq(pred, gold):
    return 1.0 if pred.strip().upper() == gold.strip().upper() else 0.0

def score_text(pred, gold, metric):
    if metric == 'BLEU-4': return compute_bleu(4, gold, pred)
    if metric == 'BERTScore': return compute_bertscore(pred, gold)
    if metric == 'RougeL': return compute_rouge_l(pred, gold)
    if metric == 'Meteor': return compute_meteor(pred, gold)
    if metric == 'F1': return compute_f1(pred, gold)
    if metric == 'MAP': return compute_map(pred, gold)

Common pitfalls

• Temperature is fixed to ~0 for all evaluations, which may penalize models that perform better with sampling or higher creativity.
• Metrics like BLEU-4 and BERTScore are reported for both QA and Summarization, but reference texts for QA are not explicitly defined (e.g., exact match vs. fuzzy matching).
• MainframeBench size and split details are not provided in the text, making reproducibility difficult without accessing the GitHub repository.

Evidence (verbatim from paper)

We evaluate these LLMs using zero-shot prompting and fixing the temperature hyperparameter to approximately 0, leading to more exploitation of the model’s current knowledge. Metrics for Multiple Choice Question task: Because it involves the direct model to select a single answer from the provided options (A, B, C, D), it is considered a classification task. We use Accuracy to report the performance of methods on multiple-choice questions. Metrics for Question Answering and COBOL Summarization task: We use various metrics in NLP, including MAP, F1-Score, BertScore, RougeL, Meteor, BLEU-4, as the evaluation metrics for these tasks.

Citation

@misc{dau2024xmainframe,
  title={XMainframe: A Large Language Model for Mainframe Modernization},
  author={Dau et al. (2024)},
  year={2024},
  note={arXiv:2408.04660}
}

• arXiv: 2408.04660