cowWhySo/permission-command-dataset

Permission Command Dataset

Synthetic and augmented command-line examples for training a local permission classifier that helps gate command execution in an application runtime.

Status, 30 August 2026

The two command models trained on this corpus were degenerate and have been withdrawn. Each emitted the same score for every input, and the metrics shipped beside them report ROC AUC 0.5000, which is what a constant scores rather than a measurement of skill.

They were command_hazard_model.onnx and obfuscation_model.onnx in cowWhySo/permission-gate-onnx. Both remain at the tag v0-degenerate-models in that repository, and its card carries the full measurement.

The tell is in the published numbers. Every average_precision equals its class base rate to the last digit, and that happens only when every score is tied:

Model and split	Reported AP	Base rate
hazard, validation	0.5142857142857142	18/35
hazard, test	0.325	13/40
obfuscation, validation	0.42857142857142855	15/35
obfuscation, test	0.525	21/40

The cause is feature scaling in the training notebooks. Raw len(s) and raw character counts were appended to two L2-normalized hashing blocks, so one column carried magnitude in the thousands beside two blocks whose entire norm was 1.0. Training then used alpha=1e-6 under learning_rate="optimal", which derives the step size from alpha. That asked for a very large step and no regularization at once, the weights diverged, and the sigmoid saturated.

Bounding those features and raising alpha to 1e-4 fixes it. The same pipeline then reaches test ROC AUC 0.9972 for the hazard model and 0.9774 for the obfuscation model.

The two notebooks confirm this between them, without meaning to. Both train with alpha=1e-6. The ONNX notebook's build_command_sparse_features appends the lexical block and produces the same ROC AUC of 0.5000, so its exported command_hazard_model.onnx and obfuscation_model.onnx are dead too. Its build_pi_sparse_features omits that block and returns hstack([x_char, x_word]) alone, and that model scores 0.9811.

One column is the only difference between those two outcomes. That is a controlled comparison sitting in the published artifacts already.

The fixed numbers still need a caveat

Held out by generator, the hazard model falls from 0.9972 to 0.7010. 181 of the 246 training rows come from a single model, so much of the in-distribution score is the classifier recognizing that model's phrasing. Report the held-out number alongside the pooled one.

See the permission-command-corpus card for the split statistics and for what this corpus does not answer, and the prompt-injection-watch-dataset card for a larger source confound found in the sibling corpus.

Labels

• actual_label: ground-truth good versus bad
• surface_label: whether the command appears benign or suspicious at a glance
• obfuscation_label: plain versus obfuscated
• obfuscation_type: specific obfuscation family
• risk_class: broad behavioral category

Splits

Split	Rows	Positive rate
train	246	0.504
validation	35	0.514
test	40	0.325

Caveats

• Contains synthetic examples
• Contains LLM-generated examples, 181 of 246 training rows from one generator
• The test split is 40 rows, which is too few to separate models that are close
• Must be reviewed and expanded with real-world false positives and false negatives before production use