Build Success Predictions from TravisTorrent and GHALogs, 2013-2023

175,706 automated software builds from two open-source CI/CD platforms spanning 10 years, collected by Lalit Narayan Mishra. The data supports a study on temporal data leakage in machine learning models for predicting build success or failure. Models using only pre-build features achieved 82.73% accuracy on TravisTorrent (2013-2017) and 83.30% on GHALogs (2023).

Use Cases

• Training build outcome classifiers based on static project metadata and build history.
• Benchmarking machine learning models for software engineering tasks using the provided accuracy figures.
• Studying the impact of project maturity and organizational factors on build success prediction.
• Developing tools to detect temporal data leakage in software prediction tasks.

Strengths

• Dataset contains 175,706 builds from two distinct platforms, TravisTorrent (100,000 builds) and GHALogs (75,706 workflows).
• Covers a 10-year time range, from 2013 to 2023, allowing for temporal analysis.
• Models built on the data achieved specific, reported accuracy scores (82.73% and 83.30%) using only legitimate pre-build features.
• Data is associated with a clear methodology for identifying and preventing temporal data leakage.

Limitations

• Row count for the actual dataset file is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset file is very small (5.5 KB), suggesting it may contain summary or metadata rather than the full raw build records.

Provenance

Source

TravisTorrent and GHALogs open-source CI/CD platforms.

Collection Method

Build logs and metadata were likely collected and processed for the described study.

Time Range

2013 to 2023

Freshness

Last updated 2026-05-27 17:39:37; freshness should be verified.