Replace toxicity model with multi-label classifier

DESIGN.md

@@ -6,12 +6,7 @@ Toxic Content Filter is a browser extension that automatically identifies and hi
 
 The extension performs all inference locally. Social-media content is never sent to an external classification service.
 
-The initial classifier produces a simple binary classification:
+The classifier produces independent toxicity-category probabilities. Users enable the categories they want filtered and configure one shared threshold; content is filtered when any enabled category reaches it.
-
-* **Toxic**
-* **Non-toxic**
-
-The model should also expose its confidence/probability, allowing the user to configure the threshold at which content is filtered.
 
 ## Goals
 
@@ -27,16 +22,7 @@ The extension should:
 
 ## Classification Model
 
-The initial implementation will use a small transformer model trained specifically for toxicity classification.
+The implementation uses `wagesj45/multilabel-toxic-comment-classifier`, a multilingual ModernBERT classifier distributed under Apache-2.0. It produces probability scores for toxicity, severe toxicity, obscene, threat, insult, identity attack, and sexual explicit content.
-
-A candidate is Citizen Lab's multilingual DistilBERT toxicity classifier, which produces:
-
-```text
-toxic
-not_toxic
-```
-
-along with classification probabilities.
 
 The model should be converted to ONNX and quantized as aggressively as practical while retaining acceptable classification accuracy.
 
@@ -45,7 +31,7 @@ The deployed extension does not require Python or a native companion application
 Conceptually:
 
 ```text
-DistilBERT toxicity model
+ModernBERT multi-label toxicity model
           |
           v
         ONNX

README.md

@@ -35,18 +35,18 @@ python3 -m pip install -r tools/requirements-model.txt
 
 # Inspect architecture, labels, tokenizer, and resolved source revision.
 python3 tools/convert_model.py inspect \
-  --model wagesj45/toxic-comment-classifier
+  --model wagesj45/multilabel-toxic-comment-classifier
 
 # Use the immutable commit printed by inspection for a release artifact.
 python3 tools/convert_model.py prepare \
-  --model wagesj45/toxic-comment-classifier \
+  --model wagesj45/multilabel-toxic-comment-classifier \
   --revision <resolved-huggingface-commit>
 python3 tools/convert_model.py validate
 ```
 
-Preparation writes the Transformers.js-compatible files and `model-manifest.json` under `public/models/toxicity/`. The manifest records the source revision, Apache 2.0 license, toxic/non-toxic label indices, maximum sequence length, and int8 quantization format. Source PyTorch/safetensors weights are never copied into the extension.
+Preparation writes the Transformers.js-compatible files and `model-manifest.json` under `public/models/toxicity/`. The manifest records the source revision, Apache 2.0 license, ordered multi-label category mapping, 512-token maximum sequence length, and int8 quantization format. Source PyTorch/safetensors weights are never copied into the extension.
 
-If the model uses generic labels such as `LABEL_0` and `LABEL_1`, pass `--toxic-index` and `--non-toxic-index` to `prepare`; the command refuses to guess an ambiguous mapping.
+The bundled [multi-label ModernBERT model](https://huggingface.co/wagesj45/multilabel-toxic-comment-classifier) produces independent scores for toxicity, severe toxicity, obscene, threat, insult, identity attack, and sexual explicit content. Preparation refuses models that do not expose exactly this label set.
 
 VibeGuard injects a lightweight content script on ordinary web pages, but reads and processes content only when a validated site definition matches the page. The extension includes a bundled definition collection for the official Mastodon web interface, Threads, old Reddit, and X. The Threads definition covers both `threads.com` and legacy `threads.net` URLs using semantic and data-attribute selectors; the Reddit definition targets `old.reddit.com`, while the X definition covers both `x.com` and legacy `twitter.com` URLs. The default subscription follows the repository's raw JSON collection and refreshes weekly, while the bundled copy remains available offline. Additional JSON subscriptions, local overrides, and disabled definitions are managed in the options page. Definitions are declarative selectors and URL patterns, never executable code.
 
@@ -71,7 +71,7 @@ Inference runs locally through ONNX Runtime's WASM backend using the bundled qua
 
 ## Options page
 
-Open the extension’s options page to change the threshold, filtering mode, and score display. The Site definitions section manages JSON subscriptions, which refresh automatically once per week and can also be refreshed manually. Local overrides can be imported from or exported to JSON and are validated before they are saved. A failed or invalid subscription update leaves the last-known-good definitions active.
+Open the extension’s options page to change the shared category threshold, enabled categories, filtering mode, and score display. A post is filtered when any enabled category reaches the threshold, and a collapsed placeholder identifies the highest-scoring matching category. The Site definitions section manages JSON subscriptions, which refresh automatically once per week and can also be refreshed manually. Local overrides can be imported from or exported to JSON and are validated before they are saved. A failed or invalid subscription update leaves the last-known-good definitions active.
 
 The options page uses the locally bundled [Bulma CSS](https://bulma.io/) v1.0.3 stylesheet. No options-page styling or runtime dependency is loaded from a CDN.
 
@@ -88,7 +88,7 @@ VibeGuard requests the following permissions:
 
 ## Model and third-party licenses
 
-The packaged toxicity model is distributed under the Apache License 2.0. Its source revision, label mapping, sequence length, and quantization details are recorded in `public/models/toxicity/model-manifest.json`.
+The packaged toxicity model is distributed under the Apache License 2.0. Its source revision, multi-label mapping, sequence length, and quantization details are recorded in `public/models/toxicity/model-manifest.json`.
 
 VibeGuard also bundles [Bulma CSS v1.0.3](https://github.com/jgthms/bulma), which is distributed under the MIT License. The project’s own code is licensed under the terms in [LICENSE](LICENSE).
 

public/models/toxicity/LICENSE

@@ -0,0 +1,197 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or Object
+      form, made available under the License, as indicated by a copyright
+      notice that is included in or attached to the work (an example is
+      provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the Work
+      or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You meet
+      the following conditions:
+
+      (a) You must give any other recipients of the Work or Derivative Works
+          a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works that
+          You distribute, all copyright, patent, trademark, and attribution
+          notices from the Source form of the Work, excluding those notices
+          that do not pertain to any part of the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

public/models/toxicity/config.json

@@ -1,37 +1,98 @@
 {
-  "activation": "gelu",
   "architectures": [
-    "DistilBertForSequenceClassification"
+    "ModernBertForSequenceClassification"
   ],
-  "attention_dropout": 0.1,
+  "attention_bias": false,
-  "bos_token_id": null,
+  "attention_dropout": 0.0,
-  "dim": 768,
+  "bos_token_id": 2,
-  "dropout": 0.1,
+  "classifier_activation": "gelu",
+  "classifier_bias": false,
+  "classifier_dropout": 0.0,
+  "classifier_pooling": "mean",
+  "cls_token_id": 1,
+  "decoder_bias": true,
+  "deterministic_flash_attn": false,
   "dtype": "float32",
-  "eos_token_id": null,
+  "embedding_dropout": 0.0,
-  "hidden_dim": 3072,
+  "eos_token_id": 1,
+  "global_attn_every_n_layers": 3,
+  "gradient_checkpointing": false,
+  "hidden_activation": "gelu",
+  "hidden_size": 384,
   "id2label": {
-    "0": "not_toxic",
+    "0": "toxicity",
-    "1": "toxic"
+    "1": "severe_toxicity",
+    "2": "obscene",
+    "3": "threat",
+    "4": "insult",
+    "5": "identity_attack",
+    "6": "sexual_explicit"
   },
+  "initializer_cutoff_factor": 2.0,
   "initializer_range": 0.02,
+  "intermediate_size": 1152,
   "label2id": {
-    "not_toxic": 0,
+    "identity_attack": 5,
-    "toxic": 1
+    "insult": 4,
+    "obscene": 2,
+    "severe_toxicity": 1,
+    "sexual_explicit": 6,
+    "threat": 3,
+    "toxicity": 0
   },
-  "max_position_embeddings": 512,
+  "layer_norm_eps": 1e-05,
-  "model_type": "distilbert",
+  "layer_types": [
-  "n_heads": 12,
+    "full_attention",
-  "n_layers": 6,
+    "sliding_attention",
-  "output_past": true,
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "local_attention": 128,
+  "mask_token_id": 4,
+  "max_position_embeddings": 8192,
+  "mlp_bias": false,
+  "mlp_dropout": 0.0,
+  "model_type": "modernbert",
+  "norm_bias": false,
+  "norm_eps": 1e-05,
+  "num_attention_heads": 6,
+  "num_hidden_layers": 22,
   "pad_token_id": 0,
-  "problem_type": "single_label_classification",
+  "position_embedding_type": "sans_pos",
-  "qa_dropout": 0.1,
+  "problem_type": "multi_label_classification",
-  "seq_classif_dropout": 0.2,
+  "rope_parameters": {
-  "sinusoidal_pos_embds": false,
+    "full_attention": {
-  "tie_weights_": true,
+      "rope_theta": 160000.0,
+      "rope_type": "default"
+    },
+    "sliding_attention": {
+      "rope_theta": 160000.0,
+      "rope_type": "default"
+    }
+  },
+  "sep_token_id": 1,
+  "sparse_pred_ignore_index": -100,
+  "sparse_prediction": false,
   "tie_word_embeddings": true,
   "transformers_version": "5.15.1",
   "use_cache": false,
-  "vocab_size": 119547
+  "vocab_size": 256000
 }

public/models/toxicity/model-manifest.json

@@ -1,15 +1,19 @@
 {
-  "source": "wagesj45/toxic-comment-classifier",
+  "source": "wagesj45/multilabel-toxic-comment-classifier",
-  "revision": "a7d2df2ead42f0bce00b330939574a02266772f5",
+  "revision": "d90fd72e603240957e36d6afc8829d6988decff6",
-  "architecture": "DistilBertForSequenceClassification",
+  "architecture": "ModernBertForSequenceClassification",
   "labels": {
-    "toxic": 1,
-    "nonToxic": 0,
     "names": {
-      "not_toxic": 0,
+      "toxicity": 0,
-      "toxic": 1
+      "severe_toxicity": 1,
+      "obscene": 2,
+      "threat": 3,
+      "insult": 4,
+      "identity_attack": 5,
+      "sexual_explicit": 6
     }
   },
+  "scoring": "multi-label-sigmoid",
   "maxLength": 512,
   "quantization": "int8-dynamic",
   "runtime": "onnxruntime-web-wasm",

public/models/toxicity/tokenizer_config.json

@@ -1,15 +1,25 @@
 {
   "backend": "tokenizers",
-  "cls_token": "[CLS]",
+  "bos_token": "<bos>",
-  "do_lower_case": false,
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "<bos>",
+  "eos_token": "<eos>",
+  "extra_special_tokens": [
+    "<start_of_turn>",
+    "<end_of_turn>"
+  ],
   "is_local": false,
   "local_files_only": false,
-  "mask_token": "[MASK]",
+  "mask_token": "<mask>",
-  "model_max_length": 512,
+  "model_input_names": [
-  "pad_token": "[PAD]",
+    "input_ids",
-  "sep_token": "[SEP]",
+    "attention_mask"
-  "strip_accents": null,
+  ],
-  "tokenize_chinese_chars": true,
+  "model_max_length": 8192,
-  "tokenizer_class": "BertTokenizer",
+  "pad_token": "<pad>",
-  "unk_token": "[UNK]"
+  "padding_side": "right",
+  "sep_token": "<eos>",
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": "<unk>"
 }

public/options.html

@@ -17,12 +17,18 @@
         <p class="subtitle">Configure local toxicity filtering for supported sites.</p>
         <form id="settings">
           <div class="field">
-            <label class="label" for="threshold">Toxicity threshold: <output id="threshold-value">80%</output></label>
+            <label class="label" for="threshold">Category threshold: <output id="threshold-value">80%</output></label>
             <div class="control">
               <input class="slider is-fullwidth" id="threshold" type="range" min="0" max="1" step=".01">
             </div>
           </div>
 
+          <fieldset class="field">
+            <legend class="label">Categories to filter</legend>
+            <p class="help mb-2">A post is filtered when any selected category reaches the threshold.</p>
+            <div id="enabled-labels" class="content"></div>
+          </fieldset>
+
           <div class="field">
             <label class="label" for="filter-mode">Filtering mode</label>
             <div class="control">
@@ -36,7 +42,7 @@
           </div>
 
           <div class="field">
-            <label class="checkbox"><input id="show-score" type="checkbox"> Show toxicity score</label>
+            <label class="checkbox"><input id="show-score" type="checkbox"> Show highest matching category score</label>
           </div>
 
           <fieldset class="field">

src/content/filter.ts

@@ -1,4 +1,4 @@
-import type { InferenceResult, Settings } from "../shared/types";
+import { TOXICITY_LABELS, type InferenceResult, type Settings, type ToxicityLabel } from "../shared/types";
 
 const HIDDEN = "data-vibeguard-hidden";
 const ORIGINAL_DISPLAY = "data-vibeguard-original-display";
@@ -11,7 +11,8 @@ export interface FilterOptions {
 }
 
 export function applyResult(element: Element, result: InferenceResult, settings: Settings, options: FilterOptions = {}): void {
-  const shouldFilter = result.label === "toxic" && result.probability >= settings.threshold;
+  const selected = selectFilterLabel(result, settings);
+  const shouldFilter = selected !== undefined;
   const postId = options.postId ?? result.id;
   if (!shouldFilter) { restore(element, postId); return; }
 
@@ -31,7 +32,7 @@ export function applyResult(element: Element, result: InferenceResult, settings:
     element.insertAdjacentElement("afterend", placeholder);
   }
   placeholder.className = "vibeguard-placeholder";
-  const message = `Content hidden as toxic${settings.showScore ? ` (${Math.round(result.probability * 100)}%)` : ""}`;
+  const message = `Content hidden as ${formatLabel(selected!.label)}${settings.showScore ? ` (${Math.round(selected!.probability * 100)}%)` : ""}`;
   const button = document.createElement("button");
   button.type = "button";
   button.textContent = "Show";
@@ -42,6 +43,17 @@ export function applyResult(element: Element, result: InferenceResult, settings:
   placeholder.replaceChildren(message, " ", button);
 }
 
+export function selectFilterLabel(result: InferenceResult, settings: Settings): { label: ToxicityLabel; probability: number } | undefined {
+  return TOXICITY_LABELS
+    .filter((label) => settings.enabledLabels.includes(label) && result.scores[label] >= settings.threshold)
+    .map((label) => ({ label, probability: result.scores[label] }))
+    .sort((a, b) => b.probability - a.probability)[0];
+}
+
+function formatLabel(label: ToxicityLabel): string {
+  return label.replaceAll("_", " ");
+}
+
 export function hasAppliedFilter(element: Element, postId: string, settings: Settings): boolean {
   const html = element as HTMLElement;
   return element.getAttribute(HIDDEN) === "true" && html.style.display === "none"

src/content/main.ts

@@ -1,6 +1,6 @@
 import { effectiveDefinitions } from "./definitions";
 import { DefinitionEngine, hasPotentialDefinitionForUrl, selectDefinition, selectDefinitionForUrl } from "./definition-engine";
-import { applyResult, hasAppliedFilter, removeOrphanedPlaceholders, restore } from "./filter";
+import { applyResult, hasAppliedFilter, removeOrphanedPlaceholders, restore, selectFilterLabel } from "./filter";
 import { loadSettings } from "../shared/settings";
 import { hashText } from "../shared/hash";
 import { priorityFor } from "../inference/queue";
@@ -95,7 +95,7 @@ function runDefinition(settings: Settings, sequence: number): () => void {
 
     for (const [id, result] of results) {
       const elements = elementsById.get(id) ?? [];
-      const shouldFilter = result.label === "toxic" && result.probability >= settings.threshold && !revealedPostIds.has(id);
+      const shouldFilter = selectFilterLabel(result, settings) !== undefined && !revealedPostIds.has(id);
       for (const element of elements) {
         if (!shouldFilter) restore(element, id);
         else if (!hasAppliedFilter(element, id, settings)) {

src/inference/classifier.ts

@@ -1,6 +1,6 @@
 import { env, pipeline, type TextClassificationPipeline } from "@huggingface/transformers";
 import { hashText } from "../shared/hash";
-import type { InferenceRequest, InferenceResult } from "../shared/types";
+import { TOXICITY_LABELS, type InferenceRequest, type InferenceResult, type ToxicityLabel } from "../shared/types";
 import { loadModelManifest, type ModelManifest } from "./model-metadata";
 
 let classifier: TextClassificationPipeline | undefined;
@@ -17,22 +17,21 @@ export async function classify(requests: InferenceRequest[], modelBaseUrl: strin
   classifier ??= await loadClassifier(modelBaseUrl);
   const invoke = classifier as unknown as (texts: string[], options: Record<string, unknown>) => Promise<Array<Array<{ label: string; score: number }> | { label: string; score: number }>>;
   const startedAt = performance.now();
-  // Transformers.js defaults `top_k` to 1. Request both scores because the
+  const outputs = await invoke(requests.map((request) => request.text), { top_k: TOXICITY_LABELS.length, max_length: modelManifest.maxLength });
-  // filtering threshold is based on the toxic class probability, including
-  // when the non-toxic class is the model's highest-confidence prediction.
-  const outputs = await invoke(requests.map((request) => request.text), { top_k: 2, max_length: modelManifest.maxLength });
   const modelDurationMs = performance.now() - startedAt;
   return requests.map((request, index) => {
     const output = Array.isArray(outputs[index]) ? outputs[index] : [outputs[index]];
-    const toxic = output.find((item) => resolveOutputIndex(item?.label, modelManifest) === modelManifest.labels.toxic);
+    const scores = Object.fromEntries(TOXICITY_LABELS.map((label) => {
-    if (!toxic) throw new Error(`Classifier output did not contain toxic label index ${modelManifest.labels.toxic}`);
+      const outputIndex = modelManifest.labels.names[label];
-    const probability = Math.max(0, Math.min(1, Number(toxic?.score ?? 0)));
+      const score = output.find((item) => resolveOutputIndex(item?.label, modelManifest) === outputIndex)?.score;
+      if (score === undefined) throw new Error(`Classifier output did not contain ${label} label index ${outputIndex}`);
+      return [label, Math.max(0, Math.min(1, Number(score)))];
+    })) as Record<ToxicityLabel, number>;
     return {
       requestId: request.requestId,
       id: request.id,
       textHash: hashText(request.text),
-      label: probability >= 0.5 ? "toxic" : "not_toxic",
+      scores,
-      probability,
       navigationId: request.navigationId,
       modelRevision: modelManifest.revision,
       modelDurationMs
@@ -49,9 +48,7 @@ async function loadClassifier(modelBaseUrl: string): Promise<TextClassificationP
 export function resolveOutputIndex(label: string | undefined, metadata: ModelManifest): number | undefined {
   if (!label) return undefined;
   const normalized = label.toLowerCase().replace(/[\s-]+/g, "_");
-  if (normalized === "toxic") return metadata.labels.toxic;
+  const configured = metadata.labels.names[normalized as ToxicityLabel] ?? metadata.labels.names[label as ToxicityLabel];
-  if (normalized === "not_toxic" || normalized === "non_toxic" || normalized === "non-toxic") return metadata.labels.nonToxic;
-  const configured = metadata.labels.names[normalized] ?? metadata.labels.names[label];
   if (configured !== undefined) return configured;
   const match = normalized.match(/^label_(\d+)$/);
   return match ? Number(match[1]) : undefined;

src/inference/model-metadata.ts

@@ -1,12 +1,13 @@
+import { TOXICITY_LABELS, type ToxicityLabel } from "../shared/types";
+
 export interface ModelManifest {
   source: string;
   revision: string;
   architecture: string;
   labels: {
-    toxic: number;
+    names: Record<ToxicityLabel, number>;
-    nonToxic: number;
-    names: Record<string, number>;
   };
+  scoring: "multi-label-sigmoid";
   maxLength: number;
   quantization: string;
   runtime: string;
@@ -26,7 +27,9 @@ export function loadModelManifest(modelBaseUrl: string): Promise<ModelManifest>
 
 export function validateManifest(value: ModelManifest): ModelManifest {
   if (!value || typeof value.source !== "string" || typeof value.revision !== "string") throw new Error("Invalid VibeGuard model manifest");
-  if (!Number.isInteger(value.labels?.toxic) || !Number.isInteger(value.labels?.nonToxic) || value.labels.toxic === value.labels.nonToxic || !value.labels?.names) throw new Error("Model manifest has no binary label mapping");
+  if (value.scoring !== "multi-label-sigmoid" || !value.labels?.names) throw new Error("Model manifest has no multi-label scoring contract");
+  const labelIndexes = TOXICITY_LABELS.map((label) => value.labels.names[label]);
+  if (labelIndexes.some((index) => !Number.isInteger(index) || index < 0) || new Set(labelIndexes).size !== TOXICITY_LABELS.length || Object.keys(value.labels.names).length !== TOXICITY_LABELS.length) throw new Error("Model manifest has an invalid label mapping");
   if (!Number.isInteger(value.maxLength) || value.maxLength < 8) throw new Error("Model manifest has an invalid maximum length");
   return value;
 }

src/options/main.ts

@@ -1,6 +1,6 @@
 import { loadSettings, saveDefinitionConfiguration, saveSettings } from "../shared/settings";
 import { validateDefinitionConfiguration } from "../shared/site-definitions";
-import type { FilterMode, Settings } from "../shared/types";
+import { TOXICITY_LABELS, type FilterMode, type Settings, type ToxicityLabel } from "../shared/types";
 import "./style.css";
 
 const form = document.querySelector<HTMLFormElement>("#settings");
@@ -8,6 +8,7 @@ const threshold = document.querySelector<HTMLInputElement>("#threshold");
 const thresholdValue = document.querySelector<HTMLElement>("#threshold-value");
 const mode = document.querySelector<HTMLSelectElement>("#filter-mode");
 const showScore = document.querySelector<HTMLInputElement>("#show-score");
+const enabledLabels = document.querySelector<HTMLElement>("#enabled-labels");
 const status = document.querySelector<HTMLElement>("#status");
 const definitions = document.querySelector<HTMLTextAreaElement>("#definitions");
 const importButton = document.querySelector<HTMLButtonElement>("#import-definitions");
@@ -24,6 +25,7 @@ void loadSettings().then((settings) => {
   threshold!.value = String(settings.threshold);
   mode!.value = settings.filterMode;
   showScore!.checked = settings.showScore;
+  renderEnabledLabels(settings);
   definitions!.value = JSON.stringify(definitionConfiguration(settings), null, 2);
   renderSubscriptions(settings);
   updateThresholdLabel();
@@ -60,7 +62,7 @@ form?.addEventListener("submit", async (event) => {
   try {
     const configuration = parseDefinitionConfiguration();
     await saveDefinitionConfiguration(configuration);
-    await saveSettings({ threshold: Number(threshold?.value), filterMode: mode?.value as FilterMode, showScore: showScore?.checked });
+    currentSettings = await saveSettings({ threshold: Number(threshold?.value), filterMode: mode?.value as FilterMode, showScore: showScore?.checked, enabledLabels: selectedLabels() });
     showStatus("Saved. Reload matching pages to apply definition changes.");
   } catch (error) { showStatus(error instanceof Error ? error.message : String(error), true); }
 });
@@ -96,6 +98,30 @@ function updateThresholdLabel(): void {
   if (thresholdValue && threshold) thresholdValue.textContent = `${Math.round(Number(threshold.value) * 100)}%`;
 }
 
+function renderEnabledLabels(settings: Settings): void {
+  if (!enabledLabels) return;
+  enabledLabels.replaceChildren(...TOXICITY_LABELS.map((label) => {
+    const input = document.createElement("input");
+    input.type = "checkbox";
+    input.name = "enabled-label";
+    input.value = label;
+    input.checked = settings.enabledLabels.includes(label);
+    const text = document.createTextNode(` ${formatLabel(label)}`);
+    const wrapper = document.createElement("label");
+    wrapper.className = "checkbox mr-4";
+    wrapper.append(input, text);
+    return wrapper;
+  }));
+}
+
+function selectedLabels(): ToxicityLabel[] {
+  return TOXICITY_LABELS.filter((label) => document.querySelector<HTMLInputElement>(`input[name="enabled-label"][value="${label}"]`)?.checked);
+}
+
+function formatLabel(label: ToxicityLabel): string {
+  return label.replaceAll("_", " ");
+}
+
 function definitionConfiguration(settings?: Settings): { customDefinitions: Settings["customDefinitions"]; disabledDefinitionIds: string[]; subscriptions: Settings["subscriptions"]; subscriptionCollections: Settings["subscriptionCollections"] } {
   return { customDefinitions: settings?.customDefinitions ?? [], disabledDefinitionIds: settings?.disabledDefinitionIds ?? [], subscriptions: settings?.subscriptions ?? [], subscriptionCollections: settings?.subscriptionCollections ?? [] };
 }

src/shared/settings.ts

@@ -1,4 +1,4 @@
-import { DEFAULT_SETTINGS, type Settings, type SiteDefinition } from "./types";
+import { DEFAULT_SETTINGS, TOXICITY_LABELS, type Settings, type SiteDefinition, type ToxicityLabel } from "./types";
 import { validateDefinitionConfiguration } from "./site-definitions";
 import type { DefinitionSubscription } from "./types";
 
@@ -59,6 +59,7 @@ export function mergeSettings(input?: StoredSettings): Settings {
     ...DEFAULT_SETTINGS,
     ...current,
     threshold: clamp(Number(current.threshold ?? DEFAULT_SETTINGS.threshold), 0, 1),
+    enabledLabels: normalizeEnabledLabels(current.enabledLabels),
     customDefinitions: definitionConfiguration.valid ? definitionConfiguration.value.customDefinitions : DEFAULT_SETTINGS.customDefinitions,
     disabledDefinitionIds: definitionConfiguration.valid ? definitionConfiguration.value.disabledDefinitionIds : DEFAULT_SETTINGS.disabledDefinitionIds,
     subscriptions: definitionConfiguration.valid ? definitionConfiguration.value.subscriptions : [DEFAULT_SUBSCRIPTION],
@@ -66,6 +67,12 @@ export function mergeSettings(input?: StoredSettings): Settings {
   };
 }
 
+function normalizeEnabledLabels(value: unknown): ToxicityLabel[] {
+  if (!Array.isArray(value)) return [...TOXICITY_LABELS];
+  const known = new Set(TOXICITY_LABELS);
+  return [...new Set(value.filter((label): label is ToxicityLabel => typeof label === "string" && known.has(label as ToxicityLabel)))];
+}
+
 export function saveDefinitionConfiguration(configuration: { customDefinitions: SiteDefinition[]; disabledDefinitionIds: string[]; subscriptions?: DefinitionSubscription[]; subscriptionCollections?: Settings["subscriptionCollections"] }): Promise<Settings> {
   const validation = validateDefinitionConfiguration(configuration);
   if (!validation.valid) return Promise.reject(new Error(validation.errors.join("\n")));

src/shared/types.ts

@@ -1,5 +1,7 @@
 export type QueuePriority = 0 | 1 | 2 | 3;
 export type FilterMode = "collapse" | "hide";
+export const TOXICITY_LABELS = ["toxicity", "severe_toxicity", "obscene", "threat", "insult", "identity_attack", "sexual_explicit"] as const;
+export type ToxicityLabel = typeof TOXICITY_LABELS[number];
 
 export interface SiteDefinition {
   id: string;
@@ -61,8 +63,7 @@ export interface InferenceResult {
   requestId: string;
   id: string;
   textHash: string;
-  label: "toxic" | "not_toxic";
+  scores: Record<ToxicityLabel, number>;
-  probability: number;
   navigationId: string;
   modelRevision?: string;
   /** Duration of the model invocation that produced this result, in milliseconds. */
@@ -75,6 +76,7 @@ export type InferenceResponse =
 
 export interface Settings {
   threshold: number;
+  enabledLabels: ToxicityLabel[];
   filterMode: FilterMode;
   showScore: boolean;
   customDefinitions: SiteDefinition[];
@@ -91,6 +93,7 @@ export interface PageStatus {
 
 export const DEFAULT_SETTINGS: Settings = {
   threshold: 0.8,
+  enabledLabels: [...TOXICITY_LABELS],
   filterMode: "collapse",
   showScore: true,
   customDefinitions: [],

tests/cache.test.ts

@@ -6,8 +6,7 @@ const result = (id: string): InferenceResult => ({
   requestId: id,
   id,
   textHash: id,
-  label: "toxic",
+  scores: { toxicity: .9, severe_toxicity: 0, obscene: 0, threat: 0, insult: 0, identity_attack: 0, sexual_explicit: 0 },
-  probability: .9,
   navigationId: "nav"
 });
 

tests/filter.test.ts

@@ -1,14 +1,16 @@
 import { afterEach, describe, expect, it } from "vitest";
-import { applyResult, removeOrphanedPlaceholders, restore } from "../src/content/filter";
+import { applyResult, removeOrphanedPlaceholders, restore, selectFilterLabel } from "../src/content/filter";
 import { DEFAULT_SETTINGS } from "../src/shared/types";
 
+const scores = { toxicity: .91, severe_toxicity: .1, obscene: .3, threat: .2, insult: .8, identity_attack: .1, sexual_explicit: .1 };
+
 describe("content filtering", () => {
   afterEach(() => document.body.replaceChildren());
 
   it("collapses toxic content and restores it", () => {
     const element = document.createElement("article");
     document.body.append(element);
-    applyResult(element, { requestId: "1", id: "1", textHash: "x", label: "toxic", probability: .91, navigationId: "n" }, DEFAULT_SETTINGS);
+    applyResult(element, { requestId: "1", id: "1", textHash: "x", scores, navigationId: "n" }, DEFAULT_SETTINGS);
     expect(element.style.display).toBe("none");
     expect(document.querySelector("[data-vibeguard-placeholder]")).not.toBeNull();
     restore(element);
@@ -19,13 +21,19 @@ describe("content filtering", () => {
   it("preserves the original display value when filtering is reconciled repeatedly", () => {
     const element = document.createElement("article");
     document.body.append(element);
-    const result = { requestId: "1", id: "1", textHash: "x", label: "toxic" as const, probability: .91, navigationId: "n" };
+    const result = { requestId: "1", id: "1", textHash: "x", scores, navigationId: "n" };
     applyResult(element, result, DEFAULT_SETTINGS);
     applyResult(element, result, DEFAULT_SETTINGS);
     restore(element);
     expect(element.style.display).toBe("");
   });
 
+  it("uses the highest enabled category above the threshold", () => {
+    const result = { requestId: "1", id: "1", textHash: "x", scores, navigationId: "n" };
+    expect(selectFilterLabel(result, { ...DEFAULT_SETTINGS, enabledLabels: ["insult", "threat"], threshold: .5 })).toEqual({ label: "insult", probability: .8 });
+    expect(selectFilterLabel(result, { ...DEFAULT_SETTINGS, enabledLabels: [], threshold: .5 })).toBeUndefined();
+  });
+
   it("removes placeholders no longer paired with the current post element", () => {
     const stale = document.createElement("div");
     const orphan = document.createElement("div");

tests/inference-worker.test.ts

@@ -23,7 +23,7 @@ class FakeWorker implements InferenceWorker {
   terminate(): void { this.terminated = true; }
 
   reply(): void {
-    const event = new MessageEvent("message", { data: { results: [{ requestId: request.requestId, id: request.id, textHash: "hash", label: "not_toxic", probability: .1, navigationId: request.navigationId }] } });
+    const event = new MessageEvent("message", { data: { results: [{ requestId: request.requestId, id: request.id, textHash: "hash", scores: { toxicity: .1, severe_toxicity: 0, obscene: 0, threat: 0, insult: 0, identity_attack: 0, sexual_explicit: 0 }, navigationId: request.navigationId }] } });
     this.listeners.get("message")?.forEach((listener) => listener(event));
   }
 }

tests/model-metadata.test.ts

@@ -3,25 +3,26 @@ import { validateManifest } from "../src/inference/model-metadata";
 import { resolveOutputIndex } from "../src/inference/classifier";
 
 const manifest = {
-  source: "wagesj45/toxic-comment-classifier",
+  source: "wagesj45/multilabel-toxic-comment-classifier",
   revision: "abc123",
-  architecture: "DistilBertForSequenceClassification",
+  architecture: "ModernBertForSequenceClassification",
-  labels: { toxic: 1, nonToxic: 0, names: { toxic: 1, non_toxic: 0 } },
+  labels: { names: { toxicity: 0, severe_toxicity: 1, obscene: 2, threat: 3, insult: 4, identity_attack: 5, sexual_explicit: 6 } },
+  scoring: "multi-label-sigmoid" as const,
   maxLength: 512,
   quantization: "int8-dynamic",
   runtime: "onnxruntime-web-wasm"
 };
 
 describe("model contract", () => {
-  it("accepts a binary manifest and resolves common output labels", () => {
+  it("accepts a multi-label manifest and resolves output labels", () => {
     expect(validateManifest(manifest)).toEqual(manifest);
     expect(resolveOutputIndex("LABEL_1", manifest)).toBe(1);
-    expect(resolveOutputIndex("not-toxic", manifest)).toBe(0);
+    expect(resolveOutputIndex("identity-attack", manifest)).toBe(5);
-    expect(resolveOutputIndex("toxic", manifest)).toBe(1);
+    expect(resolveOutputIndex("toxicity", manifest)).toBe(0);
   });
 
   it("rejects incomplete metadata", () => {
-    expect(() => validateManifest({ ...manifest, labels: { toxic: 1, nonToxic: 1, names: {} } })).toThrow();
+    expect(() => validateManifest({ ...manifest, labels: { names: { ...manifest.labels.names, threat: 1 } } })).toThrow();
     expect(() => validateManifest({ ...manifest, maxLength: 0 })).toThrow();
     expect(() => validateManifest({ ...manifest, labels: undefined as unknown as typeof manifest.labels })).toThrow();
   });

tests/queue.test.ts

@@ -11,11 +11,11 @@ describe("InferenceQueue", () => {
     let calls = 0;
     const queue = new InferenceQueue(async (requests) => {
       calls += 1;
-      return requests.map((item) => ({ requestId: item.requestId, id: item.id, textHash: "hash", label: "toxic" as const, probability: .9, navigationId: item.navigationId }));
+      return requests.map((item) => ({ requestId: item.requestId, id: item.id, textHash: "hash", scores: { toxicity: .9, severe_toxicity: 0, obscene: 0, threat: 0, insult: 0, identity_attack: 0, sexual_explicit: 0 }, navigationId: item.navigationId }));
     }, undefined, 10, 8, 0);
     const first = await queue.enqueue(request("a", 0));
     const second = await queue.enqueue({ ...request("b", 0), text: "text a" });
-    expect(first.probability).toBe(.9);
+    expect(first.scores.toxicity).toBe(.9);
     expect(second.requestId).toBe("b");
     expect(calls).toBe(1);
     expect(queue.cacheSize).toBe(1);

tests/settings.test.ts

@@ -24,4 +24,9 @@ describe("legacy settings", () => {
     expect(mergeSettings({ cpuThreads: 99 })).not.toHaveProperty("cpuThreads");
     expect(mergeSettings({ inferenceDevice: "webgpu" } as never)).not.toHaveProperty("inferenceDevice");
   });
+
+  it("defaults missing categories and removes invalid stored labels", () => {
+    expect(mergeSettings({}).enabledLabels).toEqual(DEFAULT_SETTINGS.enabledLabels);
+    expect(mergeSettings({ enabledLabels: ["threat", "threat", "unknown"] as never }).enabledLabels).toEqual(["threat"]);
+  });
 });

tools/convert_model.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Prepare a pinned Hugging Face binary classifier for VibeGuard."""
+"""Prepare a pinned Hugging Face multi-label classifier for VibeGuard."""
 from __future__ import annotations
 
 import argparse
@@ -11,8 +11,10 @@ import tempfile
 from pathlib import Path
 from typing import Any
 
-DEFAULT_MODEL = "wagesj45/toxic-comment-classifier"
+DEFAULT_MODEL = "wagesj45/multilabel-toxic-comment-classifier"
 DEFAULT_OUTPUT = Path("public/models/toxicity")
+DEFAULT_MAX_LENGTH = 512
+REQUIRED_LABELS = ("toxicity", "severe_toxicity", "obscene", "threat", "insult", "identity_attack", "sexual_explicit")
 
 
 def make_parser() -> argparse.ArgumentParser:
@@ -21,9 +23,7 @@ def make_parser() -> argparse.ArgumentParser:
     command.add_argument("--model", default=DEFAULT_MODEL)
     command.add_argument("--revision", help="Immutable HF commit; required by prepare")
     command.add_argument("--output", type=Path, default=DEFAULT_OUTPUT)
-    command.add_argument("--toxic-index", type=int)
+    command.add_argument("--max-length", type=int, default=DEFAULT_MAX_LENGTH)
-    command.add_argument("--non-toxic-index", type=int)
-    command.add_argument("--max-length", type=int)
     return command
 
 
@@ -62,14 +62,11 @@ def inspect_source(source: Path) -> dict[str, Any]:
     }
 
 
-def resolve_labels(details: dict[str, Any], toxic_override: int | None, non_toxic_override: int | None) -> tuple[int, int, dict[str, int]]:
+def resolve_labels(details: dict[str, Any]) -> dict[str, int]:
-    names = {str(index): value.lower().replace("-", "_").replace(" ", "_") for index, value in details["id2label"].items()}
+    indexed = {value.lower().replace("-", "_").replace(" ", "_"): int(index) for index, value in details["id2label"].items()}
-    indexed = {name: int(index) for index, name in names.items()}
+    if set(indexed) != set(REQUIRED_LABELS) or len(indexed) != len(REQUIRED_LABELS):
-    toxic = toxic_override if toxic_override is not None else next((index for name, index in indexed.items() if name in ("toxic", "toxicity")), None)
+        raise ValueError(f"Expected exactly these labels: {', '.join(REQUIRED_LABELS)}")
-    non_toxic = non_toxic_override if non_toxic_override is not None else next((index for name, index in indexed.items() if name in ("non_toxic", "nontoxic", "clean", "not_toxic")), None)
+    return {label: indexed[label] for label in REQUIRED_LABELS}
-    if toxic is None or non_toxic is None or toxic == non_toxic:
-        raise ValueError("Could not resolve distinct labels; pass --toxic-index and --non-toxic-index")
-    return toxic, non_toxic, indexed
 
 
 def prepare(args: argparse.Namespace) -> None:
@@ -78,9 +75,10 @@ def prepare(args: argparse.Namespace) -> None:
     license_id = str(source_metadata.get("license") or "").lower()
     if license_id != "apache-2.0": raise SystemExit(f"Expected Apache-2.0 model metadata, found license={source_metadata.get('license')!r}")
     details = inspect_source(source)
-    if details["numLabels"] != 2: raise SystemExit(f"Expected a binary classifier, found {details['numLabels']}")
+    if details["numLabels"] != len(REQUIRED_LABELS): raise SystemExit(f"Expected {len(REQUIRED_LABELS)} labels, found {details['numLabels']}")
-    toxic, non_toxic, names = resolve_labels(details, args.toxic_index, args.non_toxic_index)
+    names = resolve_labels(details)
-    max_length = args.max_length or details["maxLength"] or 512
+    max_length = args.max_length
+    if max_length > (details["maxLength"] or max_length): raise SystemExit(f"Requested max length {max_length} exceeds model maximum {details['maxLength']}")
     with tempfile.TemporaryDirectory(prefix="vibeguard-model-") as temporary:
         export_dir = Path(temporary) / "onnx"
         float_model = export_dir / "model.onnx"
@@ -88,12 +86,12 @@ def prepare(args: argparse.Namespace) -> None:
         quantized = export_dir / "model_quantized.onnx"
         quantize(float_model, quantized)
         args.output.mkdir(parents=True, exist_ok=True)
-        for name in ("tokenizer.json", "tokenizer_config.json", "special_tokens_map.json", "vocab.txt", "merges.txt"):
+        for name in ("tokenizer.json", "tokenizer_config.json", "special_tokens_map.json", "vocab.txt", "merges.txt", "LICENSE"):
             if (source / name).exists(): shutil.copy2(source / name, args.output / name)
         shutil.copy2(source / "config.json", args.output / "config.json")
         (args.output / "onnx").mkdir(exist_ok=True)
         shutil.copy2(quantized, args.output / "onnx" / "model_quantized.onnx")
-        manifest = {"source": args.model, "revision": source_metadata["revision"], "architecture": details["architecture"], "labels": {"toxic": toxic, "nonToxic": non_toxic, "names": names}, "maxLength": max_length, "quantization": "int8-dynamic", "runtime": "onnxruntime-web-wasm", "license": "Apache-2.0"}
+        manifest = {"source": args.model, "revision": source_metadata["revision"], "architecture": details["architecture"], "labels": {"names": names}, "scoring": "multi-label-sigmoid", "maxLength": max_length, "quantization": "int8-dynamic", "runtime": "onnxruntime-web-wasm", "license": "Apache-2.0"}
         (args.output / "model-manifest.json").write_text(json.dumps(manifest, indent=2) + "\n")
     print(json.dumps(manifest, indent=2))
 
@@ -146,7 +144,9 @@ def validate(args: argparse.Namespace) -> None:
     model_path = args.output / "onnx" / "model_quantized.onnx"
     if not manifest_path.is_file() or not model_path.is_file(): raise SystemExit(f"Missing generated model files under {args.output}")
     manifest = json.loads(manifest_path.read_text())
-    if manifest.get("quantization") != "int8-dynamic" or str(manifest.get("license")).lower() != "apache-2.0": raise SystemExit("Unexpected model quantization or license")
+    if manifest.get("quantization") != "int8-dynamic" or str(manifest.get("license")).lower() != "apache-2.0" or manifest.get("scoring") != "multi-label-sigmoid": raise SystemExit("Unexpected model quantization, scoring, or license")
+    names = manifest.get("labels", {}).get("names", {})
+    if set(names) != set(REQUIRED_LABELS) or len(set(names.values())) != len(REQUIRED_LABELS): raise SystemExit("Unexpected model label mapping")
     try:
         import onnx
         onnx.checker.check_model(str(model_path))

tools/requirements-model.txt

@@ -1,5 +1,5 @@
-huggingface_hub>=0.27,<1
+huggingface_hub>=1.5,<2
-transformers>=4.48,<5
+transformers>=5.15.1,<6
 torch>=2.1,<3
 onnx>=1.17,<2
 onnxruntime>=1.20,<2