feat(model): add the trained artifact and its fitted preprocessing state

Weights on their own classify nothing here. They were fitted on
standardised inputs and the statistics producing that standardisation
cannot be recovered from them, so shipping one without the other would
ship something unusable.

That state is JSON rather than a pickled transformer because a pickled
estimator arrives with a fit method attached, and refitting at serving
time is the defect this pipeline exists to prevent.

model/golden_row.json

@@ -0,0 +1,110 @@
+{
+  "prediction": 1,
+  "raw": {
+    "Bank": 0.0,
+    "CharContinuationRate": null,
+    "Crypto": 0.0,
+    "DegitRatioInURL": 0.0,
+    "Domain": null,
+    "DomainLength": 25.0,
+    "DomainTitleMatchScore": null,
+    "HasCopyrightInfo": 1.0,
+    "HasDescription": 0.0,
+    "HasExternalFormSubmit": 0.0,
+    "HasFavicon": 1.0,
+    "HasHiddenFields": 1.0,
+    "HasObfuscation": 0.0,
+    "HasPasswordField": 0.0,
+    "HasSocialNet": 1.0,
+    "HasSubmitButton": 1.0,
+    "HasTitle": null,
+    "IsDomainIP": 0.0,
+    "IsHTTPS": 1.0,
+    "IsResponsive": 1.0,
+    "LargestLineLength": null,
+    "LetterRatioInURL": 0.562,
+    "LineOfCode": null,
+    "NoOfAmpersandInURL": null,
+    "NoOfCSS": 20.0,
+    "NoOfDegitsInURL": null,
+    "NoOfEmptyRef": 14.0,
+    "NoOfEqualsInURL": 0.0,
+    "NoOfExternalRef": null,
+    "NoOfImage": null,
+    "NoOfJS": 27.0,
+    "NoOfLettersInURL": 18.0,
+    "NoOfObfuscatedChar": 0.0,
+    "NoOfOtherSpecialCharsInURL": 2.0,
+    "NoOfPopup": null,
+    "NoOfQMarkInURL": 0.0,
+    "NoOfSelfRedirect": 0.0,
+    "NoOfSelfRef": 2.0,
+    "NoOfSubDomain": null,
+    "NoOfURLRedirect": null,
+    "NoOfiFrame": 0.0,
+    "ObfuscationRatio": null,
+    "Pay": 0.0,
+    "Robots": 1.0,
+    "SpacialCharRatioInURL": null,
+    "TLD": "uk",
+    "TLDLegitimateProb": 0.028555,
+    "TLDLength": 2.0,
+    "Title": "citroencarclborg",
+    "URL": "https://www.citroencarclub.org.uk",
+    "URLCharProb": 0.061898722,
+    "URLLength": null,
+    "URLTitleMatchScore": null
+  },
+  "score": 0.0,
+  "vector": [
+    0.9458264112472534,
+    1.0816572904586792,
+    0.0,
+    -0.5169346332550049,
+    -1.0161705017089844,
+    0.5501726865768433,
+    -1.451346755027771,
+    2.004185438156128,
+    0.0,
+    0.0,
+    0.0,
+    0.18957722187042236,
+    0.1772976815700531,
+    -0.19320730865001678,
+    -0.20799125730991364,
+    0.0,
+    0.0,
+    0.0,
+    -0.14323538541793823,
+    1.1197621822357178,
+    1.0,
+    -0.23217405378818512,
+    -0.2989865839481354,
+    1.0,
+    -1.7388476133346558,
+    -1.8001145124435425,
+    1.0,
+    1.0,
+    1.0,
+    0.0,
+    0.0,
+    0.0,
+    -0.21410752832889557,
+    -0.4983856976032257,
+    0.0,
+    1.0,
+    1.0,
+    1.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    1.0,
+    -0.24477894604206085,
+    1.5860761404037476,
+    1.2351645231246948,
+    -0.9924272298812866,
+    2.4717586040496826,
+    -0.2637692391872406
+  ]
+}

model/knn_sklearn.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a34564d384384b7708c1561ff0f5d3ce2863d117865afad7ea5e119b300aa06
+size 2040804

model/manifest.json

@@ -0,0 +1,48 @@
+{
+  "artifact_schema_version": "v1",
+  "demoted_features": [
+    "Bank",
+    "HasExternalFormSubmit",
+    "HasFavicon",
+    "HasHiddenFields",
+    "HasSubmitButton",
+    "LargestLineLength",
+    "NoOfExternalRef",
+    "NoOfPopup",
+    "NoOfSelfRef",
+    "NoOfiFrame",
+    "Pay",
+    "Robots"
+  ],
+  "family": "knn",
+  "feature_count": 49,
+  "files": {
+    "fitted_stats.json": {
+      "bytes": 676974,
+      "sha256": "9d8d3804ec079bef9f47d4a63abe365d1a69830a5b7348d53359d00e0cfd5e65"
+    },
+    "golden_row.json": {
+      "bytes": 2344,
+      "sha256": "17c9bdda53a0567ecac0969f3b1f3a391589a40728117e3f55c42394b86d1732"
+    },
+    "knn_sklearn.joblib": {
+      "bytes": 2040804,
+      "sha256": "7a34564d384384b7708c1561ff0f5d3ce2863d117865afad7ea5e119b300aa06"
+    },
+    "metrics.json": {
+      "bytes": 2929,
+      "sha256": "29dae9342898d4e816e80e04da7b3660bad84d22ae97bd3b21c952885cbc3fe8"
+    }
+  },
+  "is_scratch": false,
+  "knn_k": 20,
+  "knn_reference_rows": 10000,
+  "library_version": "1.0.0",
+  "model_key": "knn_sklearn",
+  "model_name": "KNN (scikit-learn)",
+  "parameter_count": 500002,
+  "profile": "corrected",
+  "random_state": 42,
+  "source_git_sha": "b9fbbe0054ebffe4caabcc833f3c42a8dbc72001",
+  "test_size": 0.2
+}

model/metrics.json

@@ -0,0 +1,82 @@
+{
+  "artifact_schema_version": "v1",
+  "baseline_accuracy": 0.9248,
+  "baseline_note": "A constant 'legitimate' predictor scores this. Read every accuracy against it: on its own the number cannot tell a working detector from a constant.",
+  "n_train": 112323,
+  "n_validation": 28081,
+  "positive_class": "phishing (label 0)",
+  "profiles": {
+    "corrected": {
+      "knn_k": 20,
+      "leaky": false,
+      "models": {
+        "knn_sklearn": {
+          "accuracy": 0.9807343043338912,
+          "average_precision_phishing": 0.9674091116816042,
+          "baseline_accuracy": 0.9248,
+          "confusion_matrix": [
+            [
+              1643,
+              509
+            ],
+            [
+              32,
+              25897
+            ]
+          ],
+          "family": "knn",
+          "is_scratch": false,
+          "legitimate_f1": 0.9896627495939619,
+          "legitimate_precision": 0.9807240778610922,
+          "legitimate_recall": 0.9987658606193837,
+          "lift_over_baseline": 0.05593430433389124,
+          "name": "KNN (scikit-learn)",
+          "phishing_f1": 0.8586360073164359,
+          "phishing_precision": 0.9808955223880597,
+          "phishing_recall": 0.7634758364312267,
+          "predict_seconds": 1.3011771459860029,
+          "support_legitimate": 25929,
+          "support_phishing": 2152
+        }
+      },
+      "note": "Canonical. Serving-time statistics are the training-time statistics.",
+      "parity_vs_sibling_implementation": 0.00035611267405006946
+    },
+    "legacy": {
+      "knn_k": 6,
+      "leaky": true,
+      "models": {
+        "knn_sklearn": {
+          "accuracy": 0.9862896620490723,
+          "average_precision_phishing": 0.9533724979602588,
+          "baseline_accuracy": 0.9248,
+          "confusion_matrix": [
+            [
+              1837,
+              315
+            ],
+            [
+              70,
+              25859
+            ]
+          ],
+          "family": "knn",
+          "is_scratch": false,
+          "legitimate_f1": 0.9926107901656336,
+          "legitimate_precision": 0.9879651562619394,
+          "legitimate_recall": 0.9973003201049019,
+          "lift_over_baseline": 0.06148966204907236,
+          "name": "KNN (scikit-learn)",
+          "phishing_f1": 0.9051490514905149,
+          "phishing_precision": 0.9632931305715784,
+          "phishing_recall": 0.8536245353159851,
+          "predict_seconds": 0.5451889800024219,
+          "support_legitimate": 25929,
+          "support_phishing": 2152
+        }
+      },
+      "note": "Reconstruction of the original's configuration, including standardising each split by its own mean and standard deviation. Reported for provenance only. The scaling behind these numbers does not exist at serving time and cannot be reconstructed from a single URL, so they describe a model that cannot be deployed.",
+      "parity_vs_sibling_implementation": 0.0022791211139204445
+    }
+  }
+}

origin.json

@@ -0,0 +1,84 @@
+{
+  "byte_identical": [
+    "model/fitted_stats.json"
+  ],
+  "modules": [
+    {
+      "module": "phiusiil/__init__.py",
+      "sha256": null,
+      "source": "rewritten"
+    },
+    {
+      "edited": false,
+      "module": "phiusiil/schema.py",
+      "sha256": "f2b55eb2116e9c8584a46b86f887d4b61f72e002472d30373b3e3782f07fa641",
+      "source_sha256": "f2b55eb2116e9c8584a46b86f887d4b61f72e002472d30373b3e3782f07fa641"
+    },
+    {
+      "edited": false,
+      "module": "phiusiil/features/__init__.py",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "source_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+    },
+    {
+      "edited": false,
+      "module": "phiusiil/features/url_features.py",
+      "sha256": "f56770d221335cb6bfb65ff3e54de87f02f4db3d653a0b67db74d3fc3179e5fe",
+      "source_sha256": "f56770d221335cb6bfb65ff3e54de87f02f4db3d653a0b67db74d3fc3179e5fe"
+    },
+    {
+      "edited": false,
+      "module": "phiusiil/preprocess/__init__.py",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "source_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+    },
+    {
+      "edited": false,
+      "module": "phiusiil/preprocess/stats.py",
+      "sha256": "e7a2a69950bffdc0dccb867bb4b186a9bd5f0c85a5028987782f7254093b56c4",
+      "source_sha256": "e7a2a69950bffdc0dccb867bb4b186a9bd5f0c85a5028987782f7254093b56c4"
+    },
+    {
+      "edited": true,
+      "module": "phiusiil/preprocess/scaler.py",
+      "sha256": "e15f4153108cb3a3e3043cac3b46b7eee82e9e0008ff44252896e2f787d3a400",
+      "source_sha256": "51275d0c18b2e295f0a75451f1776a329ad23469f9ebab3cba19067ae774bed2"
+    },
+    {
+      "edited": true,
+      "module": "phiusiil/preprocess/transformer.py",
+      "sha256": "6610f94b36eb5981ac05cb74bac52777431f622fcbd234837497514a5a0fb377",
+      "source_sha256": "0c0c95dd05239ac027bb986f45818855316af00dbbea5ccf0884d1f869ae18a5"
+    },
+    {
+      "edited": false,
+      "module": "phiusiil/models/__init__.py",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "source_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+    },
+    {
+      "edited": true,
+      "module": "phiusiil/models/base.py",
+      "sha256": "aad4a59b713993800e533c2febaa82927ec13bf9c4edaf52438d92029c74fe84",
+      "source_sha256": "cfd0dbe9fa0687187a4b02bb6cfbc924fcfadc9bc9150c6a85ce5132fd735cc1"
+    },
+    {
+      "edited": true,
+      "module": "phiusiil/models/sk.py",
+      "sha256": "ba278f30245fcfae4fccddc0253149f937a0ed9732dc2b49554ab03e81df158b",
+      "source_sha256": "b247856c353b867173a70af6fdd94e5b6eee31758fa95ee8324182798b76a214"
+    }
+  ],
+  "note": "Provenance for everything copied out of the training bundle. The vendored modules were copied with the package name rewritten, so source_sha256 is the hash before that rewrite and sha256 the hash as shipped. source_artifacts records the bundle's own hashes: fitted_stats.json is byte-identical and will match, while metrics.json and golden_row.json were reduced to this one model and so will not -- the recorded hash is what they were reduced from.",
+  "reserialised": [
+    "model/knn_sklearn.joblib"
+  ],
+  "source_artifacts": {
+    "fitted_stats.json": "9d8d3804ec079bef9f47d4a63abe365d1a69830a5b7348d53359d00e0cfd5e65",
+    "golden_row.json": "2a2e64148dbb0c2d298eae5e4071ca496e642fcb98a06cc2d7cfbb3d3900ea27",
+    "metrics.json": "4ed0d21483c05158a52523ee3be4fdd323da29f4cfb7f51a71f5612b7765bc8f",
+    "models/knn_sklearn.joblib": "0c1ca136527e597c5a77dccf1dc0d6d3d61e52e3e3afa770099d48a2fc393503"
+  },
+  "source_git_sha": "b9fbbe0054ebffe4caabcc833f3c42a8dbc72001",
+  "source_repo": "https://github.com/fetiai/phishing-url-classifier"
+}