Factory-reset classifier set

What is inside

One zip. The test labels are in a separate file so you can predict blind and score afterwards.

factory-reset-classifier/
  README.txt
  LICENCE.txt
  DATA_DICTIONARY.md
  data_dictionary.csv
  TEACHING_GUIDE.md
  train.csv
  test.csv
  test_answers.csv
  train_reference.csv
  variants/imbalanced/
    train.csv
    test.csv
    test_answers.csv
    train_reference.csv
  code/baseline_stdlib.py
  code/baseline_sklearn.py
  code/evaluate.py
  generator/generate.py

• train.csv and test.csv: 1,400 and 600 rows, device_id plus the 22 features; train.csv also has the label reset_30d.
• test_answers.csv: the test labels, kept apart, with the profile, true reset age and clock columns for slicing. train_reference.csv has the same extra columns for the train rows.
• variants/imbalanced/: the same four files with about 5 percent positives.
• DATA_DICTIONARY.md, data_dictionary.csv, TEACHING_GUIDE.md, README.txt, LICENCE.txt: documentation and licences.

The baseline in code/baseline_stdlib.py is a logistic regression and a decision stump written by hand with the Python standard library, so nothing needs installing. code/baseline_sklearn.py is optional and needs scikit-learn; it was checked for syntax only. code/evaluate.py prints accuracy, precision, recall, F1, balanced accuracy and the confusion matrix, for all rows or sliced by a column of the answer file.

The 22 features

Ages are days before the acquisition time, as the phone's own clock reports them. A blank cell means the artifact was not there or could not be read. The full dictionary is in DATA_DICTIONARY.md.

The feature ideas come from the reset-indicator study (LAB-TR-0004: recovery history, setup flags, boot-reason history, clocks that print an impossible year) and the clock-anomaly study (LAB-TR-0022: simple checks that fire on most ordinary phones). How the features relate to the label is invented.

Reference answers and traps

The label is exact by construction: each phone has a known reset age, and reset_30d is 1 when it is 30 days or less. test_answers.csv holds the label for every test row, and train_reference.csv and test_answers.csv add the columns an examiner would not have: the phone's profile, the true reset age (blank if never reset), whether its clock was tampered with, and the clock offset.

Traps built in

Each one is described in the teaching guide, with the numbers. Try to find them yourself first.

What the included baseline scores

600 test rows, 175 of them positive. These are printed by code/evaluate.py. The score of the default logistic regression is not printed here, because question 4 of the challenge asks for it: run the baseline and read it off.

Two limits. The accuracy of any model here is higher than it would be on real phones, because the labels are exact, the features come from a short generator and train and test are the same process; the teaching guide lists the reasons. And a score is an indicator of how a method behaves on this table. It is consistent with the method working, and it does not prove that it works on a device.

Try it with free tools

Python 3 and a text editor are enough. A spreadsheet or a SQLite browser helps for the first look.

-- the leakage column: 0.030 for not reset, 0.929 for reset
SELECT reset_30d, COUNT(*) AS n, ROUND(AVG(reviewer_reset_note), 3) AS note_rate
FROM train GROUP BY reset_30d;

-- a clock check that fires on nearly everyone: 0.887 in both classes
SELECT reset_30d, ROUND(AVG(epoch_zero_file_time), 3) FROM train GROUP BY reset_30d;

-- one trustworthy rule: a recent, plausible wipe entry is right 266 times in 271
SELECT COUNT(*), SUM(reset_30d) FROM train
WHERE last_recovery_wipe_age_days <= 30 AND recovery_clock_implausible = 0;

The third query finds only the phones that have a recovery file, so it misses the rest of the positives. Try the challenge: Train and doubt a reset classifier (Practitioner).

Get the dataset

One zip, 182,901 bytes. It holds 17 files: the tables, the answers, the guide, the code and the generator. Check it against the hash below.

Hashes

Check every file after download. The set hash is the SHA-256 of manifest.json, which lists the size and hash of every other file and the counts quoted on this page.

Set hash (SHA-256, manifest.json)

Cite this dataset

@dataset{sectratek-lab-ds-0005,
  author    = {Addai, Daniel},
  title     = {Factory-reset classifier set},
  year      = {2026},
  version   = {1.0},
  publisher = {SectraTEK Lab},
  url       = {https://lab.sectratek.com/datasets/factory-reset-classifier/},
  note      = {DS-0005, SHA-256 of manifest.json 22621655320ddf47b86af1d9963b100976ef4032f3c630586c5c9b9ac5140911}
}

Related

• Report: Android reset indicators and their clocksThe recovery, boot-reason and setup-flag sources behind the features
• Report: Base rates of clock anomaliesWhy simple clock checks fire on most ordinary phones
• Dataset: Synthetic recovery-history setGenerated recovery files with reference answers

Datasets carry their own licence. This one is CC BY 4.0 for data and MIT for code. Articles are CC BY 4.0. All licences.