Replicate the Benchmark Inputs and Task Surface

1

Clone the public bundle

Start here to get the 17 directly shipped CSVs, the full 30-task metadata, benchmark definitions, and omitted-task recreation tooling.

2

Clone the benchmark repo

You need a separate local clone of dctabgan-benchmark-expansion if you want to recreate the omitted tasks exactly.

3

Recreate omitted tasks locally

Use the included scripts to copy and verify the 13 omitted processed CSVs into your local release clone.

Source-of-truth definitions

The full task surface is inherited from the May 11 no-DoS 30D benchmark definition, and the active operational equal-dose definition inherits that same dataset surface:

metadata/source_manifests/benchmark_definitions/
PRESERVED_RATIO_TRAIN500_TEST500_SINGLE_LABEL_NODOS_30D_BENCHMARK_DEFINITION_2026-05-11.json

metadata/source_manifests/benchmark_definitions/
PRESERVED_RATIO_TRAIN500_TEST500_SINGLE_LABEL_NODOS_30D_CLEAN_DCT_AND_GAN_DOSE_MATCHED_OPERATIONAL_BENCHMARK_DEFINITION_2026-05-20.json

Those files define the exact task list, processed dataset names, data folders, and split semantics used to construct the release.

Export benchmark-consistent train/test files

For immediate ML use, this repository includes a metadata-driven splitter that uses the mirrored benchmark definition and task counts, preserves the benchmark classwise_temporal row-order semantics, and strips obvious audit columns such as source_row_index by default.

python3 scripts/list_tasks.py --status all

python3 scripts/export_benchmark_splits.py \
  --task friday_bot \
  --output-dir ml_exports/friday_bot

python3 scripts/export_benchmark_splits.py \
  --task friday_bot \
  --output-dir ml_exports/friday_bot_encoded \
  --encoded

The optional --encoded path writes basic X_train.csv, y_train.csv, X_test.csv, and y_test.csv files. Numeric fills and categorical one-hot levels are fit on train rows only, then applied to test rows. This helper is a convenience for scikit-learn-style experiments, not a replacement for the benchmark repo's full treatment/evaluation pipeline.

Recommended replication workflow

git clone https://github.com/rayborg/dctabgan-ids-benchmark-datasets.git

# Obtain a separate local checkout of dctabgan-benchmark-expansion
# by whatever route is available to you. This site does not assume
# that benchmark worktree is publicly clonable.

cd dctabgan-ids-benchmark-datasets
shasum -a 256 -c SHA256SUMS.txt

python3 scripts/recreate_omitted_datasets.py \
  --benchmark-repo ../dctabgan-benchmark-expansion

python3 scripts/verify_omitted_datasets.py \
  --benchmark-repo ../dctabgan-benchmark-expansion

python3 scripts/export_benchmark_splits.py \
  --task friday_bot \
  --output-dir ml_exports/friday_bot

After that, your local clone contains the same 30-task cleaned dataset surface used by the benchmark definitions.

Protocol documents to read alongside the data

If your goal is exact replication rather than just dataset reuse, read these mirrored benchmark notes in addition to the dataset manifests:

Those files help explain how the benchmark worktree uses the cleaned CSVs, manifests, and method surface when running the actual modeling and evaluation scripts.

What this repo gives you vs. what the benchmark repo gives you

This dataset repo	The benchmark repo
Cleaned benchmark input CSVs	Training/evaluation code and manuscript-facing benchmark logic
Public-safe 17-task downloadable subset	Full materialized benchmark worktree
Machine-readable task metadata and provenance	Experiment scripts, modeling pipelines, and output tables
Omitted-task copy/verify scripts	Optional materialization/rebuild scripts for upstream-derived processed inputs