"Human candidate" means no selected observable coding-agent signal was found; it does not establish human authorship. The performance classifier retains six explicit API errors, which are not included among performance-labeled rows.

Layout

The data/ directory follows the five pipeline stages. metadata/ contains sanitized summaries, a manifest, and checksums. Operational checkpoints, API payloads, retries, secrets, and local paths are intentionally excluded.

The balanced-sample subset is the default. Other subsets expose the full mining, attribution, classification, filtering, and sampling artifacts without combining tables that have different schemas.

The optimization catalog subsets expose the original and study-refined RQ1 taxonomies as separate tables.

The rq1-model-labels and rq2-model-labels subsets expose the complete, validated outputs from GPT-5.6-sol, Gemini 3.1 Pro Preview, and Qwen3.8 27B as separate splits. These are model-specific labels for agreement analysis, not a human-adjudicated ground truth. Provider responses, prompts, and operational retry checkpoints are intentionally excluded.

The curated-labels subset provides one compact row per pull request with agentic attribution, task type, the derived performance indicator, strict human-filter decisions, and sampling labels. Human-filter and sampling fields are null for pull requests outside the populations evaluated by those stages.

from datasets import load_dataset

sample = load_dataset(
    "rcalvome/EMSE-perf-pr-study",
    "balanced-sample",
    token=True,
)

The Hub Dataset Viewer requires a PRO or Enterprise account for private datasets. This limitation does not affect authenticated downloads or loading.

Terms

The repository's processing code is MIT licensed. This dataset contains public GitHub metadata and derived labels; no blanket license is asserted over source repository content. Users are responsible for complying with GitHub and source repository terms.

Downloads last month: 11

Total file size:

4.53 GB