Datasets:

MaramAllah
/

NPFDataSet

license: unknown
task_categories:
  - tabular-regression
tags:
  - software-engineering
  - technical-debt
  - sonarqube
  - code-quality
  - mining-software-repositories
size_categories:
  - 100K<n<1M

Technical Debt Regression Dataset

One row per (repository_id, artifact_id, t0) -- a single Java file, at a weekly snapshot, in one of the source repositories. Built by scanning each repo's history week-by-week with SonarQube and pairing each snapshot's features with the technical debt that file has 3 weeks / 1 month / 3 months later.

Repositories covered

repository_id	Source
OFBiz	apache/ofbiz-framework
SystemML	apache/systemds
Groovy	apache/groovy
NiFi	apache/nifi
Guava	google/guava
OkHttp	square/okhttp

Each repo was scanned weekly (see t0 per row) starting 2015-01-01, for 60 weeks, on its default branch at scan time (trunk for OFBiz, main for SystemML/NiFi/OkHttp, master for Groovy/Guava).

Quick start

import pandas as pd
df = pd.read_csv("hf://datasets/<your-username>/td-regression-mart/combined_td_mart.csv")

⚠️ Before you train anything, read this

TD_3w / TD_1m / TD_3m / TD_*_status / path_at_* are LABELS, not features. Never include them in your feature matrix X -- they encode the future outcome you're trying to predict.
Filter each horizon independently. A row can have TD_3w observed while TD_3m is still UNKNOWN_OBSERVATION (that horizon's scan window hasn't been reached yet). Don't drop a row just because one horizon isn't ready -- filter TD_{h}_status == 'OBSERVED' per horizon, per model.
Use the provided partition column for splitting, don't re-shuffle randomly -- it's already a chronological, purged train/validation/test split designed to prevent a training row's future label window from overlapping a test row's T0. Drop partition == 'PURGED' rows entirely.

Labels

Column	Meaning
`TD_3w`	Technical debt (SonarQube remediation effort, minutes) on this file, 3 weeks after `t0`
`TD_1m`	Same, 4 weeks after `t0`
`TD_3m`	Same, 13 weeks after `t0`
`TD_{h}_status`	`OBSERVED` / `UNKNOWN_OBSERVATION` / `UNKNOWN_LINEAGE` / `EXCLUDED` -- check before using the target
`TD_{h}_exclusion_reason`	Why, if not `OBSERVED`
`path_at_{h}`	The file's path at that horizon (may differ from `t0` if renamed)

"Technical debt" = sum of remediation-effort minutes across every SonarQube issue (bug + vulnerability + code smell) active (OPEN/CONFIRMED/REOPENED) on that file at that point in time.

Row eligibility / split columns

Column	Meaning
`eligible_t0`	Whether this row is usable as a T0 at all (production `.java` file + full trailing history available)
`exclusion_reason_t0`	Why not, if `eligible_t0` is false
`lineage_confidence`	`HIGH`/`MEDIUM`/`AMBIGUOUS`/`UNSUPPORTED` -- confidence the file's identity was tracked correctly across renames/copies
`split_family`	`CHRONOLOGICAL_POOLED`, or `REPOSITORY_HELD_OUT` for any repo passed as a full holdout
`partition`	`TRAIN` / `VALIDATION` / `TEST` / `PURGED` -- drop `PURGED` rows
`max_git_feature_ts` / `max_sonar_feature_ts` / `max_feature_ts`	Leakage-audit timestamps: the latest info-date any feature in this row could see. Should never exceed `t0`.

Identity columns

origin_id, repository_id, artifact_id, lineage_id, path_at_t0, t0 (the file's real historical commit date), commit_hash_t0, sonar_analysis_key_t0, language.

Features

Git / process (trailing exact-day windows ending at t0): ncloc_t0, file_age_days, commits_{30d,90d,180d}, lines_added/deleted_{30d,90d,180d}, churn_{30d,90d,180d}, authors_{30d,90d,180d}, days_since_last_change, commit_acceleration, churn_acceleration, commit_burstiness_180d, churn_burstiness_180d.

Legacy 3-week-bucket git features (independent second measurement of similar signals): total_lines_g3, commits_g3, churn_g3, churn_rate_g3, commit_acceleration_g3, churn_acceleration_g3, commit_burstiness_g3, churn_burstiness_g3.

SonarQube snapshot at t0: finding_count_t0, finding_density_t0, bug/vulnerability/code_smell_remediation_effort_t0, technical_debt_minutes_t0, technical_debt_density_t0, {blocker,critical,major,minor,info}_finding_count_t0, {code_smell,bug,vulnerability}_count_t0, complexity_t0, cognitive_complexity_t0, duplicated_lines_density_t0, comment_lines_density_t0.

Trend + lifecycle (rolling 3w/1m/3m/6m windows, blank until full trailing history exists): finding_density_slope/volatility_*, technical_debt_density_slope/volatility_*, complexity_slope_*, cognitive_complexity_slope_*, duplication_slope_*, finding_arrivals_*, finding_closures_*.

Known limitations

identity_ambiguous_count / identity_excluded_count are always 0 (not yet implemented).
file_age_days undercounts true age for files older than (earliest scanned week − 180 days).
Lineage (rename/copy) tracking is heuristic, not a certified production system -- treat lineage_confidence != HIGH with caution for horizon-dependent targets.
Not every feature computed during pipeline construction made it into this table (e.g. ownership_entropy_180d, an alternate issue-survival label). Ask if you need one of these -- they exist in intermediate pipeline outputs and can be added.

Citation / provenance

Built with a custom SonarQube-based mining pipeline: weekly-snapshot scanning

git-window feature extraction + Sonar trend/lifecycle features, assembled into a chronologically-split regression mart. Questions about how a specific column was computed -> ask the person who generated this dataset.