|
Download README.md from STRICT-APT/strict-score-record: direct link, hf CLI and curl.
- Browser
- Download file 3.8 kB
-
https://huggingface.co/datasets/STRICT-APT/strict-score-record/resolve/main/README.md
- Command line
-
hf download hf://datasets/STRICT-APT/strict-score-record/README.md
-
curl -L -o README.md https://huggingface.co/datasets/STRICT-APT/strict-score-record/resolve/main/README.md
3.8 kB
| license: cc-by-4.0 | |
| pretty_name: STRICT score record | |
| tags: | |
| - intrusion-detection | |
| - provenance-graphs | |
| - evaluation | |
| # STRICT score record (strict_score_record_v4) | |
| Historical per-node, per-member score files for "Where the Numbers Come From: Auditing Evaluation in | |
| Provenance-Based Intrusion Detection". Each file is a PIDSMaker evaluation output (`scores_model_epoch_<k>.pkl`, written | |
| by `torch.save` despite the extension, with `pred_scores`, `y_preds`, `y_truth`, `nodes`, `node2attacks`) at its original relative path, so every | |
| manifest verifies unchanged from this directory: | |
| sha256sum -c manifests/MANIFEST_final.sha256 | |
| `manifests/MANIFEST_<rule>.sha256` bind the frozen member record under the three checkpoint rules (final epoch, | |
| minimum validation loss, best ADP); `manifests/*_SOURCE_INPUTS.sha256` bind the eight earlier records of the | |
| preregistered confirmations; `manifests/VALIDATION_INPUTS.sha256` binds the numeric validation losses of the four | |
| headline configurations, which the replay of their checkpoint and threshold choices reads. `record/` holds the frozen member record: `roster_<rule>.csv` (configuration, | |
| dataset, system, member count), `source_audit_<rule>.csv` (member, epoch, score file), `replacements.csv` | |
| (retrained members with the replaced and replacing files), `nested.csv` (same-flag configurations absorbed by a | |
| larger one). `loader.py` prints the contents of one file. Files: 5315 distinct paths. | |
| `manifests/VALIDATION_ALL.sha256` binds the numeric validation losses of every member and rule-selected epoch of the | |
| frozen record (`artifacts/triage/validation-all-20260927/validation/<configuration>_<member>_<epoch>.npz`, key `losses`), | |
| together with `epochs.csv` (every evaluated epoch's test ADP and logged validation loss), `selected.csv` (the declared | |
| epoch beside the epoch the declared selector chooses from `epochs.csv`) and `consistency.csv` (whether the saved alerts | |
| equal the test scores above the maximum validation loss, where that rule applies); `manifests/VALIDATION_ALL_INPUTS.sha256` | |
| names the score files those arrays belong to. `manifests/W2V_PAIRS.sha256` binds the identical-input word2vec reindexer | |
| pairs: `artifacts/triage/w2v-identical-20260927/<host>_s<seed>/<attempt>/<released or corrected>/final_scores.pt` holds the | |
| same keys as a score file, beside the receipts `result.json`, `started.json` (environment and bound inputs), | |
| `final_replay.json`, `execution.json` and `initial_equality.json`. Attempt directories are `attempt0` (realized | |
| dropout 0.0, seeds 0 to 9), `dropout03_attempt0` | |
| (realized dropout 0.3, seeds 0 to 4) and `repeat0` (the released member retrained on identical inputs and initial state). | |
| `record/w2v_identical_pairs/` holds their summary tables. The prepared inputs and the raw edge losses are not published. | |
| The Zenodo record (DOI 10.5281/zenodo.22930594, all versions) archives this README, the manifests, the member record, the loader | |
| and the headline validation arrays as `strict_record_metadata.tar.zst` in numbered parts, joined with | |
| `cat strict_record_metadata.tar.zst.part* > strict_record_metadata.tar.zst`. The score files, the validation arrays of every configuration and the pair files are | |
| published here, and the manifests in either copy verify them. | |
| ## Chronological reference supplement | |
| The separately versioned [strict_reference_v1](reference/strict_reference_v1/README.md) supplement includes all 192 tuning and 60 reference members, numeric validation inputs, node and time-block scores, labels, selection records, and a standalone CPU replay for the chronological Velox/Orthrus comparison. It supports numerical recomputation, not independent raw-data retraining. Historical files above are unchanged. The supplement uses CC BY 4.0 for data/documentation and Apache 2.0 for code. | |