Update Agentic Injection Review Lite to v0.1.1

Version 0.1.1 fixes repository-root resolution in the included training script, records SHA-256 digests for every locked dataset file, and discloses within-split duplicate counts. The freshly trained classifier has exactly the same complete validation, lockbox, and diagnostic decision scores as version 0.1.0. Maintainer: Ali Zakaee (ITheEqualizer).

MODEL_CARD.template.md

@@ -144,6 +144,12 @@ questions, the model sent 6 of 3,680 benign examples to review. These results
 describe the published test samples; they do not guarantee the same behavior on
 production traffic.
 
+The NVIDIA-derived validation and test splits contain 27 and 92 repeated
+normalized texts within their respective held-out domains. Domain separation
+and paired `group_id` evaluation are preserved, but row-level confidence
+intervals should not be interpreted as if every row were statistically
+independent.
+
 `metrics.json` contains the exact counts, confidence intervals, data audit,
 candidate comparisons, and broader diagnostic results. Those diagnostics are
 substantially weaker because several datasets mix prompt injection, jailbreak,
@@ -178,6 +184,13 @@ transformations, split policy, licenses, and third-party attributions. See
 [model_manifest.json](./model_manifest.json).
 The trained artifact is CC BY-SA 3.0; the source code is MIT.
 
+Version 0.1.1 fixes project-root resolution in the included training script so
+the documented repository-root command completes release assembly. It also
+records the SHA-256 digest of every locked dataset file. The model formulation,
+threshold, complete validation and lockbox decision scores, and predictions are
+unchanged from version 0.1.0; the release improves reproducibility and
+provenance rather than classifier behavior.
+
 To reproduce the release in a clean Python 3.12 environment:
 
 ```bash

README.md

@@ -70,7 +70,7 @@ model-index:
       value: 1.0000
 ---
 
-# Agentic Injection Review Lite v0.1.0
+# Agentic Injection Review Lite v0.1.1
 
 Agentic Injection Review Lite is a lightweight English text classifier for
 screening untrusted text before an AI agent uses it. It flags language that
@@ -144,6 +144,12 @@ questions, the model sent 6 of 3,680 benign examples to review. These results
 describe the published test samples; they do not guarantee the same behavior on
 production traffic.
 
+The NVIDIA-derived validation and test splits contain 27 and 92 repeated
+normalized texts within their respective held-out domains. Domain separation
+and paired `group_id` evaluation are preserved, but row-level confidence
+intervals should not be interpreted as if every row were statistically
+independent.
+
 `metrics.json` contains the exact counts, confidence intervals, data audit,
 candidate comparisons, and broader diagnostic results. Those diagnostics are
 substantially weaker because several datasets mix prompt injection, jailbreak,
@@ -178,6 +184,13 @@ transformations, split policy, licenses, and third-party attributions. See
 [model_manifest.json](./model_manifest.json).
 The trained artifact is CC BY-SA 3.0; the source code is MIT.
 
+Version 0.1.1 fixes project-root resolution in the included training script so
+the documented repository-root command completes release assembly. It also
+records the SHA-256 digest of every locked dataset file. The model formulation,
+threshold, complete validation and lockbox decision scores, and predictions are
+unchanged from version 0.1.0; the release improves reproducibility and
+provenance rather than classifier behavior.
+
 To reproduce the release in a clean Python 3.12 environment:
 
 ```bash

checksums.sha256

@@ -1,16 +1,16 @@
 bd3508fbe8e739c82aaec4c3429c1fe2c1de52d32960a47fa605aaac37db40c9  ATTRIBUTION.md
 7b8d82edce8f68538edfe88ac8c796de51ad2274033b1068bf77be23466cbcdf  LICENSE
 467416515c08b60ef9ab26725c9b5274ef0c9a27d27df3c57809354585d52be9  LICENSE-CODE
-79fb37cb228418c1c6e5410d760f04630cf0abfcf2e41a7d04a836a614e37b6e  MODEL_CARD.template.md
+4f03e04cf333f140d6a601dd37b5b4ee8dd2bc8cdefb224429950c389e23232c  MODEL_CARD.template.md
 7b8d82edce8f68538edfe88ac8c796de51ad2274033b1068bf77be23466cbcdf  MODEL_LICENSE.md
-a5d120ba623943ce9f2e50ddf1e7e29997746e69fe8c563e7430764d01f9ddd7  README.md
+7fd259e91f9c9a7f07d674491d05a5318504132eb02593b6db4aa23553c9f31c  README.md
 bd3508fbe8e739c82aaec4c3429c1fe2c1de52d32960a47fa605aaac37db40c9  THIRD_PARTY_NOTICES.md
 da1b732a1f2afdcded4094e11d9c6bf466bbfd7c5f25f79aa3994d8ab088cfac  config.json
-4159a9818ea84bcd34332b2b9bb53028a0d5e12f8039703bc913332c408495f7  dataset_manifest.json
-c1b191b92f4cd99148feacf3f083ae30a483f294127edd2154a25c1c84d75f5b  detector_config.json
-4f07d87245c9bb094eb5982c7eb4a9b55869d6eca7b0fe9ab23ea54f07545041  metrics.json
-e73971cda7bba9165cffa57bcefc0109361b8ec8783381ca91a398804e6a92e6  model.skops
-0d30629e0774b3abbde844b844bcd183ff05038096c1927f98c95a825a2b5d44  model_manifest.json
+67383a37df5687583b9944e96e0efd6bfe871929dc0245ca868729cea017a2ba  dataset_manifest.json
+1a37981b44a091ef55f6888b4f790d1e0895548da7599b2491a8a312036747c1  detector_config.json
+ac70f0200b72c322a613de9bcd57e13176296d98aac88d227a5340c595180cdf  metrics.json
+fbbd30b55c64029b883771976a35e156c4e2d0eb71dff7d911d01f18b29831ea  model.skops
+942d8405fc8df57444fd073b0c0b4f0507bc4439ca4cd8fdb720efa890e6e699  model_manifest.json
 1b7a5f27ac4fcb95dbe1581d8203fbf636dbf564b6ef10647d8818e3ae5a4d26  requirements-train.txt
 6e4faa6fa81cd15034c467bf1a78aed2dcb31f91705d643314a420c7e6e69bd7  requirements.txt
-a20871de7ff9e4da291e67d9ab8b7e22f7df65426f66167f420200a189716a58  train.py
+3c2a365d1e0e6014e42414bbf13997dc68d6dba09b4f7810fcae57918d014f6e  train.py

dataset_manifest.json

@@ -5,6 +5,11 @@
       "id": "3nesdeniz/agentic-prompt-injection-boundary-pairs",
       "revision": "a5682e7573e1c7bc4b12e64d49c0dcd90ca776cf",
       "files": ["data/train.parquet", "data/validation.parquet", "data/test.parquet"],
+      "file_sha256": {
+        "data/train.parquet": "12f7cb0d4790a684574b46165f8538f898966b47bfc9da322e12806089ad5e04",
+        "data/validation.parquet": "1538875e45b8034070815003ad24b043d71c612f7728c4fd1d24416f53b74f49",
+        "data/test.parquet": "d47124d655765d88e1b3adc0dfa25b03c9959f2c99d196c47da77d8683c7e508"
+      },
       "license": "cc-by-4.0",
       "role": "train_validation_test"
     },
@@ -12,6 +17,9 @@
       "id": "nvidia/Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1",
       "revision": "d738d4f361cc38bb4d7a42b9066776dade5332f5",
       "files": ["train.jsonl"],
+      "file_sha256": {
+        "train.jsonl": "3329da17564a7eb287e2730fc7d6956e1f4fe51e8950ac4f110b3c37e78cf3b9"
+      },
       "license": "cc-by-4.0",
       "role": "derived_train_validation_test"
     },
@@ -19,6 +27,9 @@
       "id": "databricks/databricks-dolly-15k",
       "revision": "bdd27f4d94b9c1f951818a7da7fd7aeea5dbff1a",
       "files": ["databricks-dolly-15k.jsonl"],
+      "file_sha256": {
+        "databricks-dolly-15k.jsonl": "2df9083338b4abd6bceb5635764dab5d833b393b55759dffb0959b6fcbf794ec"
+      },
       "license": "cc-by-sa-3.0",
       "role": "benign_train_validation_test"
     },
@@ -26,6 +37,9 @@
       "id": "yanismiraoui/prompt_injections",
       "revision": "bd55359f2f332afc35f277ac3dd08f7111b024c9",
       "files": ["prompt_injections.csv"],
+      "file_sha256": {
+        "prompt_injections.csv": "f4843f1841fa19b980f804796a68fc72f06841775eaba2723c768c7d772aabad"
+      },
       "license": "apache-2.0",
       "role": "positive_train_validation"
     },
@@ -33,6 +47,9 @@
       "id": "jackhhao/jailbreak-classification",
       "revision": "2f2ceeb39658696fd3f462403562b6eea5306287",
       "files": ["balanced/jailbreak_dataset_test_balanced.csv"],
+      "file_sha256": {
+        "balanced/jailbreak_dataset_test_balanced.csv": "809ef0d7c82fa05c31e818e9c3ff68f769632ac6d11cc3b43254cada577869a9"
+      },
       "license": "apache-2.0-metadata-only",
       "role": "diagnostic_only"
     },
@@ -40,6 +57,9 @@
       "id": "deepset/prompt-injections",
       "revision": "4f61ecb038e9c3fb77e21034b22511b523772cdd",
       "files": ["data/test-00000-of-00001-701d16158af87368.parquet"],
+      "file_sha256": {
+        "data/test-00000-of-00001-701d16158af87368.parquet": "39ac797cabc157eeed58435a08593b2952bb6cb16fc394a2d383f447cc7b246e"
+      },
       "license": "ambiguous-apache-2.0-or-cc-by-4.0",
       "role": "diagnostic_only"
     },
@@ -47,9 +67,11 @@
       "id": "neuralchemy/prompt-injection-benign-dataset",
       "revision": "7d70432dfcf47a821612cbf9d34e9d9e3ad20e75",
       "files": ["core/test-00000-of-00001.parquet"],
+      "file_sha256": {
+        "core/test-00000-of-00001.parquet": "fb5432035fdd8453d61f6593bfe78ee2970403c5ea59959b642a6708e2bef670"
+      },
       "license": "mixed-source-aggregate",
       "role": "diagnostic_only"
     }
   ]
 }
-

detector_config.json

@@ -2,8 +2,8 @@
   "decision_threshold": -0.7598804523157342,
   "max_input_chars": 20000,
   "model_file": "model.skops",
-  "model_sha256": "e73971cda7bba9165cffa57bcefc0109361b8ec8783381ca91a398804e6a92e6",
-  "model_version": "0.1.0",
+  "model_sha256": "fbbd30b55c64029b883771976a35e156c4e2d0eb71dff7d911d01f18b29831ea",
+  "model_version": "0.1.1",
   "negative_label": "allow",
   "numpy_version": "2.5.1",
   "positive_label": "review",

metrics.json

@@ -62,7 +62,7 @@
         "threshold": -0.30457667593634397,
         "worst_fpr_limit_ratio": 0.5736137667304015
       },
-      "training_seconds": 2.53921562503092
+      "training_seconds": 2.380033750028815
     },
     "hash_sgd_alpha_1e-5": {
       "threshold_selection": {
@@ -121,7 +121,7 @@
         "threshold": -0.7598804523157342,
         "worst_fpr_limit_ratio": 0.9082217973231358
       },
-      "training_seconds": 2.2363845839863643
+      "training_seconds": 2.087675833958201
     },
     "hash_sgd_alpha_3e-5": {
       "threshold_selection": {
@@ -180,7 +180,7 @@
         "threshold": -0.09286502046714196,
         "worst_fpr_limit_ratio": 0.5736137667304015
       },
-      "training_seconds": 2.2174021669779904
+      "training_seconds": 2.097628415969666
     }
   },
   "data_audit": {
@@ -218,6 +218,19 @@
     "training_class_counts": {
       "0": 10393,
       "1": 1722
+    },
+    "within_split_normalized_duplicates": {
+      "boundary_scenario_test": 0,
+      "boundary_validation": 0,
+      "deepset_ambiguous_diagnostic": 0,
+      "dolly_general_qa_validation": 0,
+      "dolly_open_qa_test": 0,
+      "jack_jailbreak_diagnostic": 0,
+      "neuralchemy_mixed_diagnostic": 0,
+      "nvidia_domain_test": 92,
+      "nvidia_domain_validation": 27,
+      "training": 0,
+      "yanis_attack_validation": 0
     }
   },
   "diagnostics": {
@@ -303,34 +316,34 @@
   "latency": {
     "long": {
       "1": {
-        "median_us_per_row": 7389.417,
-        "p95_us_per_row": 7458.433
+        "median_us_per_row": 7133.458,
+        "p95_us_per_row": 7167.9746
       },
       "128": {
-        "median_us_per_row": 6829.7705078125,
-        "p95_us_per_row": 6915.0828109375
+        "median_us_per_row": 6721.4423828125,
+        "p95_us_per_row": 6743.34511875
       },
       "32": {
-        "median_us_per_row": 6879.05596875,
-        "p95_us_per_row": 6896.61614375
+        "median_us_per_row": 6703.6328125,
+        "p95_us_per_row": 6741.2896
       }
     },
     "short": {
       "1": {
-        "median_us_per_row": 348.833,
-        "p95_us_per_row": 363.7086
+        "median_us_per_row": 326.666,
+        "p95_us_per_row": 338.14119999999997
       },
       "128": {
-        "median_us_per_row": 28.5426484375,
-        "p95_us_per_row": 29.1750625
+        "median_us_per_row": 27.46809375,
+        "p95_us_per_row": 28.1366546875
       },
       "32": {
-        "median_us_per_row": 35.6888125,
-        "p95_us_per_row": 35.994262500000005
+        "median_us_per_row": 34.3111875,
+        "p95_us_per_row": 34.9351625
       }
     }
   },
-  "model_version": "0.1.0",
+  "model_version": "0.1.1",
   "schema_version": 1,
   "seed": 20260805,
   "selection_policy": "validation-only; tests and diagnostics are excluded",

model.skops

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e73971cda7bba9165cffa57bcefc0109361b8ec8783381ca91a398804e6a92e6
+oid sha256:fbbd30b55c64029b883771976a35e156c4e2d0eb71dff7d911d01f18b29831ea
 size 3204271

model_manifest.json

@@ -1,9 +1,9 @@
 {
-  "dataset_lock_sha256": "4159a9818ea84bcd34332b2b9bb53028a0d5e12f8039703bc913332c408495f7",
+  "dataset_lock_sha256": "67383a37df5687583b9944e96e0efd6bfe871929dc0245ca868729cea017a2ba",
   "maintainer": "Ali Zakaee (ITheEqualizer)",
-  "metrics_sha256": "4f07d87245c9bb094eb5982c7eb4a9b55869d6eca7b0fe9ab23ea54f07545041",
-  "model_sha256": "e73971cda7bba9165cffa57bcefc0109361b8ec8783381ca91a398804e6a92e6",
-  "model_version": "0.1.0",
+  "metrics_sha256": "ac70f0200b72c322a613de9bcd57e13176296d98aac88d227a5340c595180cdf",
+  "model_sha256": "fbbd30b55c64029b883771976a35e156c4e2d0eb71dff7d911d01f18b29831ea",
+  "model_version": "0.1.1",
   "schema_version": 1,
-  "training_script_sha256": "a20871de7ff9e4da291e67d9ab8b7e22f7df65426f66167f420200a189716a58"
+  "training_script_sha256": "3c2a365d1e0e6014e42414bbf13997dc68d6dba09b4f7810fcae57918d014f6e"
 }

train.py

@@ -36,7 +36,7 @@ from sklearn.metrics import (
 from sklearn.pipeline import FeatureUnion, Pipeline
 
 SEED = 20260805
-MODEL_VERSION = "0.1.0"
+MODEL_VERSION = "0.1.1"
 
 
 @dataclass(frozen=True)
@@ -93,6 +93,26 @@ def sha256(path: Path) -> str:
     return digest.hexdigest()
 
 
+def resolve_project_root(script_path: Path, cwd: Path) -> Path:
+    resolved_script = script_path.resolve()
+    candidates = [resolved_script.parent, cwd.resolve()]
+    if len(resolved_script.parents) > 2:
+        candidates.append(resolved_script.parents[2])
+    for candidate in dict.fromkeys(candidates):
+        has_card = (candidate / "MODEL_CARD.template.md").exists() or (
+            candidate / "templates" / "MODEL_CARD.md"
+        ).exists()
+        has_dataset_lock = (candidate / "dataset_manifest.json").exists() or (
+            candidate / "manifests" / "datasets.lock.json"
+        ).exists()
+        if has_card and has_dataset_lock:
+            return candidate
+    raise FileNotFoundError(
+        "could not resolve a project root containing a model-card template "
+        "and dataset manifest"
+    )
+
+
 def download(source: Source, filename: str, cache_dir: Path) -> Path:
     return Path(
         hf_hub_download(
@@ -390,6 +410,21 @@ def build_bundle(cache_dir: Path) -> DataBundle:
         **{name: len(frame) for name, frame in tests.items()},
         **{name: len(frame) for name, frame in cleaned_diagnostics.items()},
     }
+    audit["within_split_normalized_duplicates"] = {
+        "training": int(training["normalized"].duplicated().sum()),
+        **{
+            name: int(frame["normalized"].duplicated().sum())
+            for name, frame in validation.items()
+        },
+        **{
+            name: int(frame["normalized"].duplicated().sum())
+            for name, frame in tests.items()
+        },
+        **{
+            name: int(frame["normalized"].duplicated().sum())
+            for name, frame in cleaned_diagnostics.items()
+        },
+    }
     return DataBundle(
         training=training,
         validation=validation,
@@ -923,10 +958,7 @@ def main() -> None:
     parser.add_argument("--output", type=Path, default=Path("release"))
     parser.add_argument("--cache-dir", type=Path, default=Path("cache"))
     args = parser.parse_args()
-    source_root = Path(__file__).resolve().parents[2]
-    project_root = (
-        Path.cwd() if (Path.cwd() / "MODEL_CARD.template.md").exists() else source_root
-    )
+    project_root = resolve_project_root(Path(__file__), Path.cwd())
     bundle = build_bundle(args.cache_dir)
     model, metrics = train(bundle)
     write_release(project_root, args.output, model, metrics)