Publish epoch-5 Moderation 03, frozen presets and evaluation

LICENSE

@@ -0,0 +1,202 @@
+
+                                 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.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   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 2026 Alibaba Cloud
+
+   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.

NOTICE

@@ -0,0 +1 @@
+Moderation 03 uses Qwen3.5-2B-Base (Qwen Team), distributed separately under Apache-2.0.

README.md

@@ -0,0 +1,104 @@
+---
+language:
+  - en
+  - de
+  - fr
+  - es
+  - it
+  - sv
+  - fi
+  - pl
+  - cs
+  - lv
+  - zh
+  - ja
+  - ko
+  - ru
+  - uk
+  - be
+  - kk
+license: apache-2.0
+pipeline_tag: text-classification
+library_name: pytorch
+base_model: Qwen/Qwen3.5-2B-Base
+tags:
+  - moderation
+  - multilingual
+  - custom-code
+---
+
+# Moderation 03
+
+Frozen **Qwen3.5-2B-Base** text features → **254,564,363-parameter transformer head** → **11 raw category scores**.
+Supports 17 languages: en, de, fr, es, it, sv, fi, pl, cs, lv, zh, ja, ko, ru, uk, be, kk. This is the selected epoch-5 checkpoint.
+
+[Interactive Gradio demo](https://huggingface.co/spaces/ifmain/Moderation-03) · [Full evaluation report](https://huggingface.co/spaces/ifmain/Moderation-03-report) · [Model weights](https://huggingface.co/ifmain/Moderation-03) · [Source](https://github.com/ifmain/Moderation-03)
+
+## Outputs and controls
+
+The raw sigmoid scores are **not calibrated probabilities**. They remain unchanged when selecting a preset.
+Categories: `harassment`, `harassment_threatening`, `hate`, `hate_threatening`, `self_harm`, `self_harm_instructions`, `self_harm_intent`, `sexual`, `sexual_minors`, `violence`, `violence_graphic`.
+
+Choose raw scores, one of four presets (`light`, `medium`, `high`, `corporate`), or custom per-category thresholds. A `null` threshold disables a category. Enabled categories trigger when `score >= threshold`; any trigger blocks the text. High and corporate presets also apply a finite multilingual profanity lexicon. Custom thresholds can change each category independently and optionally enable that lexicon. Raw mode makes no allow/block decision.
+
+Corporate is the strictest **content-threshold** preset. The project owner also intends to restrict non-work discussions in workplace deployments, but this release has **no workplace-topic relevance detector**. A benign statement about equal rights is not a hate ground-truth label merely because an organization chooses to restrict that topic. The published evaluation retains its original content-policy labels; it does not measure a separate work-topic policy.
+
+## Run
+
+Clone the GitHub repository, install `requirements.txt`, and run `python app.py`. Head weights and the pinned Qwen backbone download automatically. CUDA is used when available; CPU is supported but slower. The first request loads the model. Hosted demo deployments use free CPU hardware by default; no paid hardware is required by this package.
+
+```python
+from moderation03 import Moderation03
+
+model = Moderation03("ifmain/Moderation-03")
+raw = model.predict("Hello, thanks for your help.", language="en")
+medium = model.predict("Hello, thanks for your help.", language="en", preset="medium")
+custom = dict.fromkeys(model.config["categories"], None)
+custom["harassment"] = 0.4
+result = model.predict("Hello, thanks for your help.", language="en", thresholds=custom)
+print(result["raw_scores"], result["policy"])
+```
+
+`MODEL_ID` may name a local model package or Hub repository. `BACKBONE_PATH` optionally points to a local Qwen snapshot for the app. `DEVICE=cpu` or `cuda` overrides automatic device selection. Local app defaults to `127.0.0.1:7860`; Spaces binds port 7860 publicly. The app does not write submitted text to a dataset. Hugging Face/Gradio hosting has its own service policies.
+
+## Architecture and provenance
+
+Backbone: `Qwen/Qwen3.5-2B-Base`, pinned revision `b1485b2fa6dfa1287294f269f5fb618e03d52d7c`.
+Head: input size 2048, width 1024, 20 transformer encoder blocks, 16 attention heads, feed-forward size 4096, masked mean pooling, 11 classifiers. Token limit: **512**, no chat template. Longer texts are truncated and the API returns `truncated=true`.
+Published `model.safetensors` is the **FP32 head**, SHA256 `4218fb73825ff2fbbdd897a4346dd67141b15fc63d1811413929dddb88d408d3`; it is not the full Qwen model. BF16 inference/autocast follows the evaluated pipeline. The rounded BF16 head export is intentionally not substituted for the evaluated weights.
+
+Training: 100,000 examples, 70% clean / 30% character-level augmentation, 17 languages; 2,000 separate validation examples. Selected checkpoint: epoch 5, validation BCE 0.0967052458. Training data derives from [ifmain/text-moderation-02-multilingual](https://huggingface.co/datasets/ifmain/text-moderation-02-multilingual). Character substitutions and inserted separators do not establish robustness to semantic attacks. Source language assignment included uncertain groups.
+
+## Calibration and evaluation
+
+Internal threshold fitting: **828 rows / 638 source families**. Internal audit: **412 rows / 268 families**. All 11 categories have positive reference examples in all 17 languages in both partitions; translations are correlated, not independent examples. Most labels are inherited teacher estimates, not independently reviewed human ground truth. Thresholds minimize group-weighted balanced error on the fit partition, with nested preset constraints. This is threshold calibration, not probability calibration; no globally optimal thresholds are claimed.
+
+Internal held-out content-policy results:
+
+| Preset | TP | FP | TN | FN | Precision | Recall |
+|---|---:|---:|---:|---:|---:|---:|
+| Light | 109 | 35 | 244 | 24 | 75.7% | 82.0% |
+| Medium | 189 | 51 | 146 | 26 | 78.8% | 87.9% |
+| High | 245 | 24 | 121 | 22 | 91.1% | 91.8% |
+| Corporate | 252 | 22 | 115 | 23 | 92.0% | 91.6% |
+
+External evaluation: [mmathys/openai-moderation-api-evaluation](https://huggingface.co/datasets/mmathys/openai-moderation-api-evaluation), revision `84e5cf3bcd6acb3dfc70b6760451645872218a3e`, **1,680 English examples**. Unknown labels are excluded, not treated as negatives. No thresholds were fitted on this dataset. Exact normalized overlaps with used groups, including all their translations: zero; arbitrary paraphrase/pretraining contamination is not excluded. 76 external texts exceed 512 tokens and were truncated.
+
+External projections onto available benchmark categories (not full four-policy ground truth; profanity and three unsupported heads excluded):
+
+| Preset | Evaluable rows | TP | FP | TN | FN |
+|---|---:|---:|---:|---:|---:|
+| Light | 800 | 202 | 210 | 323 | 65 |
+| Medium | 859 | 493 | 196 | 141 | 29 |
+| High | 859 | 504 | 220 | 117 | 18 |
+| Corporate | 859 | 505 | 222 | 115 | 17 |
+
+## Known failures
+
+An additional 265 synthetic checks cover adult sexual content, fraud assistance, disturbing imagery, hate-related requests and violence, plus benign controls, typography changes and 10 English/Russian paraphrases. They are not independently human-reviewed and do not reproduce the researchers' undisclosed benchmark. Corporate misses 29 of 36 fraud-request variants; under the evaluated content policy it also blocks 16 of 17 benign equal-rights messages. There is no dedicated illegal-activity head; broad disturbing content is not equivalent to graphic violence. Lower thresholds introduce false positives. The reported ModerationBERT-En transformed-prompt vulnerability is **not established as fixed** by this release.
+
+The detailed report contains category/language metrics and the frozen thresholds. Public evaluation outputs contain row IDs and scores, **not source text**. Training data, source prompts, credentials, hidden-state caches and optimizer states are not distributed in this repository.
+
+## License
+
+Apache-2.0 for released code and head weights. The backbone is downloaded separately under its own Apache-2.0 license. External evaluation provenance and definitions: [OpenAI moderation-api-release](https://github.com/openai/moderation-api-release).

app.py

@@ -0,0 +1,89 @@
+"""Gradio demo: raw outputs, frozen presets, and independent custom category thresholds."""
+import json
+import os
+from pathlib import Path
+from threading import Lock
+os.environ.setdefault('GRADIO_ANALYTICS_ENABLED','False')
+import gradio as gr
+from moderation03 import Moderation03,apply_policy,MODEL_ID
+from moderation_policy import LANGS,PROFILES
+from categories import CATEGORIES
+
+ROOT=Path(__file__).resolve().parent
+CALIBRATION=json.loads((ROOT/'calibration/thresholds.json').read_text(encoding='utf-8'))
+MODEL=None
+LOCK=Lock()
+LABELS={'light':'Light / Лёгкий','medium':'Medium / Средний','high':'High / Высокий','corporate':'Corporate / Корпоративный'}
+
+def settings(preset):
+    values=[]
+    for c in CATEGORIES:
+        t=CALIBRATION['thresholds'][preset][c]
+        values.extend([t is not None,round(t,6) if t is not None else .5])
+    return [preset in ['high','corporate'],*values]
+
+def analyze(text,language,mode,preset,lexical,*controls):
+    global MODEL
+    if not text or not text.strip():raise gr.Error('Enter text / Введите текст')
+    if len(text)>50000:raise gr.Error('Maximum 50,000 characters / Максимум 50 000 символов')
+    with LOCK:
+        if MODEL is None:
+            MODEL=Moderation03(os.getenv('MODEL_ID',MODEL_ID),device=os.getenv('DEVICE') or None,
+                               backbone_path=os.getenv('BACKBONE_PATH') or None)
+        result=MODEL.scores(text)
+    thresholds=None;block_profanity=False
+    if mode=='preset':
+        thresholds=CALIBRATION['thresholds'][preset];block_profanity=preset in ['high','corporate']
+    elif mode=='custom':
+        thresholds={c:float(controls[2*i+1]) if controls[2*i] else None for i,c in enumerate(CATEGORIES)}
+        block_profanity=bool(lexical)
+    policy=None if thresholds is None else apply_policy(result['raw_scores'],thresholds,
+        block_profanity=block_profanity,text=text,language=language)
+    result.update(mode=mode,preset=preset if mode=='preset' else None,language=language,policy=policy,
+        all_presets={p:apply_policy(result['raw_scores'],CALIBRATION['thresholds'][p],
+            block_profanity=p in ['high','corporate'],text=text,language=language) for p in PROFILES})
+    summary='**Raw scores / Сырые оценки**' if policy is None else (
+        '**BLOCK / Блокировать**' if policy['block'] else '**ALLOW / Пропустить**')
+    if policy and policy['reasons']:summary+='\n\nTriggered / Сработали: '+', '.join(policy['reasons'])
+    if result['truncated']:summary+='\n\n⚠ Input truncated to 512 tokens / Текст обрезан до 512 токенов.'
+    table=[[c,result['raw_scores'][c],None if thresholds is None else thresholds[c],
+            '—' if policy is None else ('BLOCK' if policy['tags'][c] else '—')] for c in CATEGORIES]
+    comparison=[[LABELS[p],result['all_presets'][p]['block'],', '.join(result['all_presets'][p]['reasons'])] for p in PROFILES]
+    return summary,table,comparison,result
+
+def build_app():
+    with gr.Blocks(title='Moderation 03',analytics_enabled=False) as demo:
+        gr.Markdown('# Moderation 03\n17 languages · 11 raw scores · 4 presets · custom thresholds')
+        with gr.Row():
+            with gr.Column():
+                text=gr.Textbox(label='Text / Текст',lines=7,max_lines=14)
+                language=gr.Dropdown(LANGS,value='ru',label='Language / Язык (for the profanity rule)')
+                mode=gr.Radio([('Raw scores / Сырые оценки','raw'),('Preset / Пресет','preset'),
+                               ('Custom thresholds / Свои пороги','custom')],value='preset',label='Mode / Режим')
+                preset=gr.Dropdown([(LABELS[p],p) for p in PROFILES],value='medium',label='Preset / Пресет')
+                button=gr.Button('Analyze / Проверить',variant='primary')
+            with gr.Column():
+                summary=gr.Markdown('Enter text and choose a mode / Введите текст и выберите режим.')
+                table=gr.Dataframe(headers=['Category','Raw score','Threshold','Tag'],datatype=['str','number','number','str'],interactive=False)
+        comparison=gr.Dataframe(headers=['Preset','Blocked','Reasons'],datatype=['str','bool','str'],interactive=False,label='All presets / Все пресеты')
+        with gr.Accordion('Custom thresholds / Индивидуальные пороги',open=False):
+            gr.Markdown('Editing these controls switches to custom mode. Disable a category with its checkbox. / Изменение включает режим своих порогов. Снимите галочку, чтобы отключить категорию.')
+            lexical=gr.Checkbox(value=False,label='Block finite profanity lexicon / Блокировать мат по словарю')
+            controls=[];defaults=settings('medium')[1:]
+            for i,c in enumerate(CATEGORIES):
+                with gr.Row():
+                    enabled=gr.Checkbox(value=defaults[2*i],label=c)
+                    threshold=gr.Slider(0,1,value=defaults[2*i+1],step=.001,label='Threshold / Порог')
+                controls.extend([enabled,threshold])
+            for control in [lexical,*controls]:control.input(lambda:'custom',outputs=mode,queue=False)
+        preset.change(settings,inputs=preset,outputs=[lexical,*controls],queue=False)
+        details=gr.JSON(label='Raw scores and decisions / Оценки и решения JSON')
+        button.click(analyze,inputs=[text,language,mode,preset,lexical,*controls],outputs=[summary,table,comparison,details],
+                     concurrency_limit=1,api_name='moderate')
+        gr.Markdown('Scores are not calibrated probabilities. Presets are content thresholds, not a work-topic classifier. Corporate does not reliably determine whether a topic belongs in a particular workplace. / Оценки не являются калиброванными вероятностями. Пресеты не определяют, относится ли тема к работе.')
+        gr.Markdown('[Model](https://huggingface.co/ifmain/Moderation-03) · [Evaluation report](https://huggingface.co/spaces/ifmain/Moderation-03-report) · [GitHub](https://github.com/ifmain/Moderation-03)\n\nEvaluation found false positives and missed fraud requests; the reported researcher vulnerability is not established as fixed. The app does not save submitted text to a dataset.')
+    return demo.queue(max_size=8)
+
+if __name__=='__main__':
+    build_app().launch(server_name=os.getenv('GRADIO_SERVER_NAME','0.0.0.0' if os.getenv('SPACE_ID') else '127.0.0.1'),
+                       server_port=int(os.getenv('PORT','7860')),share=False)

calibration/thresholds.json

@@ -0,0 +1,78 @@
+{
+  "status": "fitted_internal_only",
+  "method": "group-weighted balanced error threshold search; nested profiles",
+  "fit_rows": 828,
+  "fit_groups": 638,
+  "thresholds": {
+    "light": {
+      "harassment": null,
+      "harassment_threatening": 0.12,
+      "hate": 0.26,
+      "hate_threatening": 0.01,
+      "self_harm": null,
+      "self_harm_instructions": 0.23,
+      "self_harm_intent": null,
+      "sexual": null,
+      "sexual_minors": 0.02,
+      "violence": null,
+      "violence_graphic": 0.02
+    },
+    "medium": {
+      "harassment": 0.45,
+      "harassment_threatening": 0.03,
+      "hate": 0.060000000000000005,
+      "hate_threatening": 0.01,
+      "self_harm": 0.12,
+      "self_harm_instructions": 0.01,
+      "self_harm_intent": 0.02,
+      "sexual": 0.59,
+      "sexual_minors": 0.01,
+      "violence": 0.09999999999999999,
+      "violence_graphic": 0.02
+    },
+    "high": {
+      "harassment": 0.24000000000000002,
+      "harassment_threatening": 0.03,
+      "hate": 0.05,
+      "hate_threatening": 0.01,
+      "self_harm": 0.03,
+      "self_harm_instructions": 0.01,
+      "self_harm_intent": 0.02,
+      "sexual": 0.34,
+      "sexual_minors": 0.01,
+      "violence": 0.09999999999999999,
+      "violence_graphic": 0.02
+    },
+    "corporate": {
+      "harassment": 0.24000000000000002,
+      "harassment_threatening": 0.01,
+      "hate": 0.05,
+      "hate_threatening": 0.01,
+      "self_harm": 0.02,
+      "self_harm_instructions": 0.01,
+      "self_harm_intent": 0.01,
+      "sexual": 0.29000000000000004,
+      "sexual_minors": 0.01,
+      "violence": 0.09,
+      "violence_graphic": 0.02
+    }
+  },
+  "standard_category_thresholds": {
+    "harassment": 0.25,
+    "harassment_threatening": 0.02,
+    "hate": 0.03,
+    "hate_threatening": 0.01,
+    "self_harm": 0.01,
+    "self_harm_instructions": 0.01,
+    "self_harm_intent": 0.01,
+    "sexual": 0.11,
+    "sexual_minors": 0.01,
+    "violence": 0.09,
+    "violence_graphic": 0.02
+  },
+  "probabilities_calibrated": false,
+  "checkpoint_sha256": "4218fb73825ff2fbbdd897a4346dd67141b15fc63d1811413929dddb88d408d3",
+  "fit_ids_sha256": "852388954817cf5ef3047eae79e8a1480cc60a1311673e439a02fbb9511cca23",
+  "policy": "Reference four-profile policy plus finite profanity rule. Research intent checks scored separately.",
+  "limitations": "Weak teacher-derived and synthetic labels. No new detector heads were trained."
+}

categories.py

@@ -0,0 +1,17 @@
+import re, json
+
+CATEGORIES = 'harassment harassment_threatening hate hate_threatening self_harm self_harm_instructions self_harm_intent sexual sexual_minors violence violence_graphic'.split()
+
+def normalize(scores):
+    result = {}
+    for name, value in scores.items():
+        key = re.sub(r'[^a-z0-9]+', '_', name.lower()).strip('_')
+        if key in result and result[key] != value:
+            raise ValueError(f'Conflicting category aliases: {name}')
+        result[key] = value
+    if set(result) != set(CATEGORIES):
+        raise ValueError(f'Unexpected category set: {sorted(result)}')
+    return {name: result[name] for name in CATEGORIES}
+
+def canonical(scores):
+    return json.dumps(normalize(scores), sort_keys=True, separators=(',', ':'), allow_nan=False)

config.json

@@ -0,0 +1,46 @@
+{
+  "backbone": "Qwen/Qwen3.5-2B-Base",
+  "backbone_revision": "b1485b2fa6dfa1287294f269f5fb618e03d52d7c",
+  "input_size": 2048,
+  "hidden_size": 1024,
+  "num_layers": 20,
+  "num_heads": 16,
+  "intermediate_size": 4096,
+  "max_length": 512,
+  "dropout": 0.1,
+  "gradient_checkpointing": true,
+  "pooling": "masked_mean",
+  "categories": [
+    "harassment",
+    "harassment_threatening",
+    "hate",
+    "hate_threatening",
+    "self_harm",
+    "self_harm_instructions",
+    "self_harm_intent",
+    "sexual",
+    "sexual_minors",
+    "violence",
+    "violence_graphic"
+  ],
+  "training": {
+    "epochs": 10,
+    "batch_size": 4,
+    "gradient_accumulation": 8,
+    "learning_rate": 0.0001,
+    "weight_decay": 0.01,
+    "warmup_fraction": 0.05,
+    "max_grad_norm": 1.0,
+    "early_stopping_patience": 3,
+    "dtype": "bfloat16",
+    "seed": 20261003,
+    "loss": "BCEWithLogitsLoss with original soft scores",
+    "cuda_memory_fraction": 0.40875889995586007
+  },
+  "calibration": {
+    "status": "fitted_internal_only",
+    "file": "calibration/thresholds.json",
+    "probabilities_calibrated": false
+  },
+  "checkpoint_epoch": 5
+}

evaluation/coverage_audit.json

@@ -0,0 +1,489 @@
+{
+  "status": "coverage_expanded_ready_for_measurement",
+  "internal_rows": 1240,
+  "splits": {
+    "fit": 828,
+    "audit": 412
+  },
+  "challenge_rows": 265,
+  "category_language_positive_counts_at_reference_0_5": {
+    "fit": {
+      "en": {
+        "harassment": 10,
+        "harassment_threatening": 3,
+        "hate": 3,
+        "hate_threatening": 1,
+        "self_harm": 3,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 3,
+        "sexual": 6,
+        "sexual_minors": 4,
+        "violence": 11,
+        "violence_graphic": 5
+      },
+      "de": {
+        "harassment": 17,
+        "harassment_threatening": 5,
+        "hate": 4,
+        "hate_threatening": 3,
+        "self_harm": 2,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 2,
+        "sexual": 7,
+        "sexual_minors": 5,
+        "violence": 9,
+        "violence_graphic": 3
+      },
+      "fr": {
+        "harassment": 12,
+        "harassment_threatening": 4,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 4,
+        "self_harm_instructions": 3,
+        "self_harm_intent": 3,
+        "sexual": 10,
+        "sexual_minors": 5,
+        "violence": 12,
+        "violence_graphic": 5
+      },
+      "es": {
+        "harassment": 17,
+        "harassment_threatening": 3,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 5,
+        "self_harm_instructions": 3,
+        "self_harm_intent": 5,
+        "sexual": 7,
+        "sexual_minors": 5,
+        "violence": 8,
+        "violence_graphic": 3
+      },
+      "it": {
+        "harassment": 12,
+        "harassment_threatening": 4,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 4,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 4,
+        "sexual": 10,
+        "sexual_minors": 5,
+        "violence": 10,
+        "violence_graphic": 4
+      },
+      "sv": {
+        "harassment": 14,
+        "harassment_threatening": 4,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 4,
+        "self_harm_instructions": 3,
+        "self_harm_intent": 4,
+        "sexual": 7,
+        "sexual_minors": 4,
+        "violence": 7,
+        "violence_graphic": 3
+      },
+      "fi": {
+        "harassment": 11,
+        "harassment_threatening": 4,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 3,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 3,
+        "sexual": 9,
+        "sexual_minors": 5,
+        "violence": 10,
+        "violence_graphic": 4
+      },
+      "pl": {
+        "harassment": 14,
+        "harassment_threatening": 4,
+        "hate": 3,
+        "hate_threatening": 1,
+        "self_harm": 4,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 4,
+        "sexual": 8,
+        "sexual_minors": 5,
+        "violence": 10,
+        "violence_graphic": 4
+      },
+      "cs": {
+        "harassment": 16,
+        "harassment_threatening": 5,
+        "hate": 7,
+        "hate_threatening": 2,
+        "self_harm": 5,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 4,
+        "sexual": 7,
+        "sexual_minors": 4,
+        "violence": 7,
+        "violence_graphic": 3
+      },
+      "lv": {
+        "harassment": 17,
+        "harassment_threatening": 4,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 3,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 3,
+        "sexual": 9,
+        "sexual_minors": 4,
+        "violence": 8,
+        "violence_graphic": 4
+      },
+      "zh": {
+        "harassment": 15,
+        "harassment_threatening": 4,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 5,
+        "self_harm_instructions": 3,
+        "self_harm_intent": 4,
+        "sexual": 7,
+        "sexual_minors": 4,
+        "violence": 9,
+        "violence_graphic": 3
+      },
+      "ja": {
+        "harassment": 13,
+        "harassment_threatening": 3,
+        "hate": 5,
+        "hate_threatening": 2,
+        "self_harm": 4,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 4,
+        "sexual": 6,
+        "sexual_minors": 4,
+        "violence": 7,
+        "violence_graphic": 3
+      },
+      "ko": {
+        "harassment": 12,
+        "harassment_threatening": 2,
+        "hate": 3,
+        "hate_threatening": 1,
+        "self_harm": 4,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 4,
+        "sexual": 9,
+        "sexual_minors": 5,
+        "violence": 8,
+        "violence_graphic": 4
+      },
+      "ru": {
+        "harassment": 17,
+        "harassment_threatening": 3,
+        "hate": 4,
+        "hate_threatening": 1,
+        "self_harm": 3,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 3,
+        "sexual": 8,
+        "sexual_minors": 5,
+        "violence": 5,
+        "violence_graphic": 2
+      },
+      "uk": {
+        "harassment": 14,
+        "harassment_threatening": 3,
+        "hate": 3,
+        "hate_threatening": 1,
+        "self_harm": 4,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 4,
+        "sexual": 6,
+        "sexual_minors": 5,
+        "violence": 7,
+        "violence_graphic": 3
+      },
+      "be": {
+        "harassment": 14,
+        "harassment_threatening": 3,
+        "hate": 5,
+        "hate_threatening": 2,
+        "self_harm": 4,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 3,
+        "sexual": 12,
+        "sexual_minors": 5,
+        "violence": 9,
+        "violence_graphic": 3
+      },
+      "kk": {
+        "harassment": 12,
+        "harassment_threatening": 3,
+        "hate": 2,
+        "hate_threatening": 1,
+        "self_harm": 3,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 3,
+        "sexual": 10,
+        "sexual_minors": 5,
+        "violence": 10,
+        "violence_graphic": 5
+      }
+    },
+    "audit": {
+      "en": {
+        "harassment": 5,
+        "harassment_threatening": 4,
+        "hate": 4,
+        "hate_threatening": 3,
+        "self_harm": 3,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 3,
+        "sexual": 3,
+        "sexual_minors": 2,
+        "violence": 5,
+        "violence_graphic": 1
+      },
+      "de": {
+        "harassment": 6,
+        "harassment_threatening": 4,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 4,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 4,
+        "sexual": 3,
+        "sexual_minors": 1,
+        "violence": 5,
+        "violence_graphic": 3
+      },
+      "fr": {
+        "harassment": 9,
+        "harassment_threatening": 3,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 2,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 2,
+        "sexual": 3,
+        "sexual_minors": 2,
+        "violence": 3,
+        "violence_graphic": 1
+      },
+      "es": {
+        "harassment": 8,
+        "harassment_threatening": 4,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 2,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 2,
+        "sexual": 4,
+        "sexual_minors": 2,
+        "violence": 5,
+        "violence_graphic": 2
+      },
+      "it": {
+        "harassment": 7,
+        "harassment_threatening": 3,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 3,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 3,
+        "sexual": 5,
+        "sexual_minors": 2,
+        "violence": 4,
+        "violence_graphic": 1
+      },
+      "sv": {
+        "harassment": 9,
+        "harassment_threatening": 3,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 2,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 2,
+        "sexual": 4,
+        "sexual_minors": 2,
+        "violence": 4,
+        "violence_graphic": 2
+      },
+      "fi": {
+        "harassment": 7,
+        "harassment_threatening": 3,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 4,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 4,
+        "sexual": 4,
+        "sexual_minors": 2,
+        "violence": 5,
+        "violence_graphic": 1
+      },
+      "pl": {
+        "harassment": 7,
+        "harassment_threatening": 3,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 2,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 2,
+        "sexual": 3,
+        "sexual_minors": 1,
+        "violence": 4,
+        "violence_graphic": 1
+      },
+      "cs": {
+        "harassment": 6,
+        "harassment_threatening": 3,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 3,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 2,
+        "sexual": 3,
+        "sexual_minors": 1,
+        "violence": 4,
+        "violence_graphic": 2
+      },
+      "lv": {
+        "harassment": 7,
+        "harassment_threatening": 3,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 2,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 2,
+        "sexual": 5,
+        "sexual_minors": 2,
+        "violence": 4,
+        "violence_graphic": 1
+      },
+      "zh": {
+        "harassment": 6,
+        "harassment_threatening": 3,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 2,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 2,
+        "sexual": 4,
+        "sexual_minors": 2,
+        "violence": 6,
+        "violence_graphic": 2
+      },
+      "ja": {
+        "harassment": 6,
+        "harassment_threatening": 5,
+        "hate": 4,
+        "hate_threatening": 2,
+        "self_harm": 3,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 3,
+        "sexual": 3,
+        "sexual_minors": 1,
+        "violence": 6,
+        "violence_graphic": 2
+      },
+      "ko": {
+        "harassment": 7,
+        "harassment_threatening": 4,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 2,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 3,
+        "sexual": 5,
+        "sexual_minors": 1,
+        "violence": 4,
+        "violence_graphic": 1
+      },
+      "ru": {
+        "harassment": 6,
+        "harassment_threatening": 3,
+        "hate": 3,
+        "hate_threatening": 3,
+        "self_harm": 3,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 3,
+        "sexual": 2,
+        "sexual_minors": 1,
+        "violence": 6,
+        "violence_graphic": 2
+      },
+      "uk": {
+        "harassment": 7,
+        "harassment_threatening": 3,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 3,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 2,
+        "sexual": 3,
+        "sexual_minors": 1,
+        "violence": 4,
+        "violence_graphic": 2
+      },
+      "be": {
+        "harassment": 6,
+        "harassment_threatening": 4,
+        "hate": 3,
+        "hate_threatening": 2,
+        "self_harm": 2,
+        "self_harm_instructions": 1,
+        "self_harm_intent": 2,
+        "sexual": 2,
+        "sexual_minors": 1,
+        "violence": 7,
+        "violence_graphic": 2
+      },
+      "kk": {
+        "harassment": 9,
+        "harassment_threatening": 4,
+        "hate": 5,
+        "hate_threatening": 3,
+        "self_harm": 3,
+        "self_harm_instructions": 2,
+        "self_harm_intent": 3,
+        "sexual": 3,
+        "sexual_minors": 1,
+        "violence": 5,
+        "violence_graphic": 1
+      }
+    }
+  },
+  "all_11_categories_positive_in_all_17_languages_both_splits": true,
+  "external_rows": 1680,
+  "external_overlap_excluded_from_primary_test": 0,
+  "external_revision": "84e5cf3bcd6acb3dfc70b6760451645872218a3e",
+  "external_sha256": "ef3f7de86ee394fe337b0a566ba4b9c7db058f3caa8d6a4a0006138b663ab9a8",
+  "research_mapping": {
+    "pornography": [
+      "sexual"
+    ],
+    "illegal_activity": [],
+    "disturbing_content": [
+      "violence_graphic",
+      "self_harm"
+    ],
+    "hate": [
+      "hate",
+      "hate_threatening"
+    ],
+    "violence": [
+      "violence",
+      "violence_graphic"
+    ]
+  },
+  "limitations": [
+    "Coverage is not proof of detection. Illegal activity has no dedicated output.",
+    "Broad disturbing content is not equivalent to graphic violence.",
+    "Internal source labels are inherited teacher estimates; authored checks are not human-reviewed.",
+    "Translations share semantic groups; reported row counts are not independent sample counts.",
+    "Typography and 10 authored EN/RU semantic paraphrases do not reproduce the undisclosed researcher benchmark.",
+    "External data has unknown labels and no gold four-profile labels.",
+    "Challenge policy: fraud assistance, hateful content, praise of violence and gross imagery blocked at all levels; adult erotica allowed only at light."
+  ]
+}

evaluation/index.html

@@ -0,0 +1 @@
+<!doctype html><html lang="ru"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Moderation 03 — калибровка и тест</title><style>body{font:16px system-ui;background:#101723;color:#eef3fc;max-width:1150px;margin:36px auto;padding:0 24px}h1,h2{color:#8cc9ff}h2{margin-top:36px}table{border-collapse:collapse;width:100%;font-variant-numeric:tabular-nums}th,td{text-align:right;padding:10px;border-bottom:1px solid #354457}th:first-child{text-align:left}details{margin:14px 0}summary{cursor:pointer;color:#8cc9ff}.note{padding:20px;background:#302719}.scroll{overflow:auto}a{color:#8cc9ff}p{line-height:1.6}</style><h1>Moderation 03 · эпоха 5</h1><p class="note">Этот отчёт проверяет исходную политику контента. Владелец также предполагает запрет нерабочих тем в корпоративных чатах; отдельный детектор релевантности рабочей теме здесь не реализован и не оценивался. Запрет темы организацией не меняет её сырой тег на hate.</p><p>Пороги подобраны только на 828 внутренних примерах. Отложенная проверка: 412 примеров. Внешний тест: 1 680. Веса модели фиксированы.</p><p class="note">Это калибровка порогов, не подтверждение исправления атаки исследователей. Внутренние метки преимущественно унаследованы от исходного датасета; новые проверки синтетические. Переводы связаны: число строк не равно числу независимых ситуаций. Незаконная деятельность не имеет отдельного выхода модели.</p><h2>Внутренний отложенный тест · после калибровки</h2><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>412</td><td>109</td><td>35</td><td>244</td><td>24</td><td>75.7%</td><td>82.0%</td><td>12.5%</td><td>78.7%</td></tr><tr><th>Средний</th><td>412</td><td>189</td><td>51</td><td>146</td><td>26</td><td>78.8%</td><td>87.9%</td><td>25.9%</td><td>83.1%</td></tr><tr><th>Высокий</th><td>412</td><td>245</td><td>24</td><td>121</td><td>22</td><td>91.1%</td><td>91.8%</td><td>16.6%</td><td>91.4%</td></tr><tr><th>Корпоративный</th><td>412</td><td>252</td><td>22</td><td>115</td><td>23</td><td>92.0%</td><td>91.6%</td><td>16.1%</td><td>91.8%</td></tr></table></div><details><summary>До калибровки</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>412</td><td>20</td><td>0</td><td>279</td><td>113</td><td>100.0%</td><td>15.0%</td><td>0.0%</td><td>26.1%</td></tr><tr><th>Средний</th><td>412</td><td>118</td><td>19</td><td>178</td><td>97</td><td>86.1%</td><td>54.9%</td><td>9.6%</td><td>67.0%</td></tr><tr><th>Высокий</th><td>412</td><td>208</td><td>7</td><td>138</td><td>59</td><td>96.7%</td><td>77.9%</td><td>4.8%</td><td>86.3%</td></tr><tr><th>Корпоративный</th><td>412</td><td>246</td><td>19</td><td>118</td><td>29</td><td>92.8%</td><td>89.5%</td><td>13.9%</td><td>91.1%</td></tr></table></div></details><h2>Теги по уровням · внутренний тест</h2><details><summary>Лёгкий</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>harassment_threatening</th><td>412</td><td>17</td><td>22</td><td>351</td><td>22</td><td>43.6%</td><td>43.6%</td><td>5.9%</td><td>43.6%</td></tr><tr><th>hate</th><td>412</td><td>39</td><td>7</td><td>350</td><td>16</td><td>84.8%</td><td>70.9%</td><td>2.0%</td><td>77.2%</td></tr><tr><th>hate_threatening</th><td>412</td><td>26</td><td>10</td><td>365</td><td>11</td><td>72.2%</td><td>70.3%</td><td>2.7%</td><td>71.2%</td></tr><tr><th>self_harm_instructions</th><td>412</td><td>4</td><td>1</td><td>388</td><td>19</td><td>80.0%</td><td>17.4%</td><td>0.3%</td><td>28.6%</td></tr><tr><th>sexual_minors</th><td>412</td><td>23</td><td>18</td><td>369</td><td>2</td><td>56.1%</td><td>92.0%</td><td>4.7%</td><td>69.7%</td></tr><tr><th>violence_graphic</th><td>412</td><td>22</td><td>46</td><td>340</td><td>4</td><td>32.4%</td><td>84.6%</td><td>11.9%</td><td>46.8%</td></tr></table></div></details><details><summary>Средний</summary><div class="scroll"><table><tr><th>Уровен�� / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>harassment</th><td>412</td><td>82</td><td>37</td><td>265</td><td>28</td><td>68.9%</td><td>74.5%</td><td>12.3%</td><td>71.6%</td></tr><tr><th>harassment_threatening</th><td>412</td><td>53</td><td>59</td><td>294</td><td>6</td><td>47.3%</td><td>89.8%</td><td>16.7%</td><td>62.0%</td></tr><tr><th>hate</th><td>412</td><td>55</td><td>40</td><td>315</td><td>2</td><td>57.9%</td><td>96.5%</td><td>11.3%</td><td>72.4%</td></tr><tr><th>hate_threatening</th><td>412</td><td>26</td><td>10</td><td>365</td><td>11</td><td>72.2%</td><td>70.3%</td><td>2.7%</td><td>71.2%</td></tr><tr><th>self_harm</th><td>412</td><td>18</td><td>5</td><td>370</td><td>19</td><td>78.3%</td><td>48.6%</td><td>1.3%</td><td>60.0%</td></tr><tr><th>self_harm_instructions</th><td>412</td><td>18</td><td>24</td><td>363</td><td>7</td><td>42.9%</td><td>72.0%</td><td>6.2%</td><td>53.7%</td></tr><tr><th>self_harm_intent</th><td>412</td><td>27</td><td>12</td><td>364</td><td>9</td><td>69.2%</td><td>75.0%</td><td>3.2%</td><td>72.0%</td></tr><tr><th>sexual</th><td>412</td><td>21</td><td>6</td><td>378</td><td>7</td><td>77.8%</td><td>75.0%</td><td>1.6%</td><td>76.4%</td></tr><tr><th>sexual_minors</th><td>412</td><td>23</td><td>26</td><td>361</td><td>2</td><td>46.9%</td><td>92.0%</td><td>6.7%</td><td>62.2%</td></tr><tr><th>violence</th><td>412</td><td>63</td><td>66</td><td>279</td><td>4</td><td>48.8%</td><td>94.0%</td><td>19.1%</td><td>64.3%</td></tr><tr><th>violence_graphic</th><td>412</td><td>23</td><td>45</td><td>339</td><td>5</td><td>33.8%</td><td>82.1%</td><td>11.7%</td><td>47.9%</td></tr></table></div></details><details><summary>Высокий</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>harassment</th><td>412</td><td>130</td><td>43</td><td>215</td><td>24</td><td>75.1%</td><td>84.4%</td><td>16.7%</td><td>79.5%</td></tr><tr><th>harassment_threatening</th><td>412</td><td>56</td><td>56</td><td>293</td><td>7</td><td>50.0%</td><td>88.9%</td><td>16.0%</td><td>64.0%</td></tr><tr><th>hate</th><td>412</td><td>62</td><td>35</td><td>311</td><td>4</td><td>63.9%</td><td>93.9%</td><td>10.1%</td><td>76.1%</td></tr><tr><th>hate_threatening</th><td>412</td><td>26</td><td>10</td><td>365</td><td>11</td><td>72.2%</td><td>70.3%</td><td>2.7%</td><td>71.2%</td></tr><tr><th>self_harm</th><td>412</td><td>31</td><td>15</td><td>354</td><td>12</td><td>67.4%</td><td>72.1%</td><td>4.1%</td><td>69.7%</td></tr><tr><th>self_harm_instructions</th><td>412</td><td>18</td><td>24</td><td>363</td><td>7</td><td>42.9%</td><td>72.0%</td><td>6.2%</td><td>53.7%</td></tr><tr><th>self_harm_intent</th><td>412</td><td>29</td><td>10</td><td>363</td><td>10</td><td>74.4%</td><td>74.4%</td><td>2.7%</td><td>74.4%</td></tr><tr><th>sexual</th><td>412</td><td>28</td><td>11</td><td>342</td><td>31</td><td>71.8%</td><td>47.5%</td><td>3.1%</td><td>57.1%</td></tr><tr><th>sexual_minors</th><td>412</td><td>23</td><td>26</td><td>361</td><td>2</td><td>46.9%</td><td>92.0%</td><td>6.7%</td><td>62.2%</td></tr><tr><th>violence</th><td>412</td><td>68</td><td>61</td><td>274</td><td>9</td><td>52.7%</td><td>88.3%</td><td>18.2%</td><td>66.0%</td></tr><tr><th>violence_graphic</th><td>412</td><td>24</td><td>44</td><td>339</td><td>5</td><td>35.3%</td><td>82.8%</td><td>11.5%</td><td>49.5%</td></tr></table></div></details><details><summary>Корпоративный</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>harassment</th><td>412</td><td>136</td><td>37</td><td>208</td><td>31</td><td>78.6%</td><td>81.4%</td><td>15.1%</td><td>80.0%</td></tr><tr><th>harassment_threatening</th><td>412</td><td>62</td><td>97</td><td>250</td><td>3</td><td>39.0%</td><td>95.4%</td><td>28.0%</td><td>55.4%</td></tr><tr><th>hate</th><td>412</td><td>70</td><td>27</td><td>310</td><td>5</td><td>72.2%</td><td>93.3%</td><td>8.0%</td><td>81.4%</td></tr><tr><th>hate_threatening</th><td>412</td><td>26</td><td>10</td><td>365</td><td>11</td><td>72.2%</td><td>70.3%</td><td>2.7%</td><td>71.2%</td></tr><tr><th>self_harm</th><td>412</td><td>35</td><td>19</td><td>348</td><td>10</td><td>64.8%</td><td>77.8%</td><td>5.2%</td><td>70.7%</td></tr><tr><th>self_harm_instructions</th><td>412</td><td>18</td><td>24</td><td>363</td><td>7</td><td>42.9%</td><td>72.0%</td><td>6.2%</td><td>53.7%</td></tr><tr><th>self_harm_intent</th><td>412</td><td>37</td><td>16</td><td>352</td><td>7</td><td>69.8%</td><td>84.1%</td><td>4.3%</td><td>76.3%</td></tr><tr><th>sexual</th><td>412</td><td>32</td><td>10</td><td>336</td><td>34</td><td>76.2%</td><td>48.5%</td><td>2.9%</td><td>59.3%</td></tr><tr><th>sexual_minors</th><td>412</td><td>24</td><td>25</td><td>361</td><td>2</td><td>49.0%</td><td>92.3%</td><td>6.5%</td><td>64.0%</td></tr><tr><th>violence</th><td>412</td><td>90</td><td>46</td><td>264</td><td>12</td><td>66.2%</td><td>88.2%</td><td>14.8%</td><td>75.6%</td></tr><tr><th>violence_graphic</th><td>412</td><td>24</td><td>44</td><td>339</td><td>5</td><td>35.3%</td><td>82.8%</td><td>11.5%</td><td>49.5%</td></tr></table></div></details><h2>Пять направлений из письма · дополнительные проверки</h2><p>Положительные запросы, безопасные контрпримеры, изменения написания; 10 переформулировок EN/RU. Это небольшой диагностический набор, не внешний бенчмарк. Эротика взрослых разрешена только на лёгком уровне; остальные четыре направления блокируются на всех уровнях.</p><h3>disturbing_content</h3><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>53</td><td>23</td><td>0</td><td>17</td><td>13</td><td>100.0%</td><td>63.9%</td><td>0.0%</td><td>78.0%</td></tr><tr><th>Средний</th><td>53</td><td>24</td><td>0</td><td>17</td><td>12</td><td>100.0%</td><td>66.7%</td><td>0.0%</td><td>80.0%</td></tr><tr><th>Высокий</th><td>53</td><td>27</td><td>1</td><td>16</td><td>9</td><td>96.4%</td><td>75.0%</td><td>5.9%</td><td>84.4%</td></tr><tr><th>Корпоративный</th><td>53</td><td>27</td><td>1</td><td>16</td><td>9</td><td>96.4%</td><td>75.0%</td><td>5.9%</td><td>84.4%</td></tr></table></div><h3>hate</h3><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>53</td><td>22</td><td>4</td><td>13</td><td>14</td><td>84.6%</td><td>61.1%</td><td>23.5%</td><td>71.0%</td></tr><tr><th>Средний</th><td>53</td><td>36</td><td>16</td><td>1</td><td>0</td><td>69.2%</td><td>100.0%</td><td>94.1%</td><td>81.8%</td></tr><tr><th>Высокий</th><td>53</td><td>36</td><td>16</td><td>1</td><td>0</td><td>69.2%</td><td>100.0%</td><td>94.1%</td><td>81.8%</td></tr><tr><th>Корпоративный</th><td>53</td><td>36</td><td>16</td><td>1</td><td>0</td><td>69.2%</td><td>100.0%</td><td>94.1%</td><td>81.8%</td></tr></table></div><h3>illegal_activity</h3><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>53</td><td>0</td><td>0</td><td>17</td><td>36</td><td>—</td><td>0.0%</td><td>0.0%</td><td>0.0%</td></tr><tr><th>Средний</th><td>53</td><td>0</td><td>0</td><td>17</td><td>36</td><td>—</td><td>0.0%</td><td>0.0%</td><td>0.0%</td></tr><tr><th>Высокий</th><td>53</td><td>2</td><td>0</td><td>17</td><td>34</td><td>100.0%</td><td>5.6%</td><td>0.0%</td><td>10.5%</td></tr><tr><th>Корпоративный</th><td>53</td><td>7</td><td>0</td><td>17</td><td>29</td><td>100.0%</td><td>19.4%</td><td>0.0%</td><td>32.6%</td></tr></table></div><h3>pornography</h3><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>53</td><td>0</td><td>32</td><td>21</td><td>0</td><td>0.0%</td><td>—</td><td>60.4%</td><td>0.0%</td></tr><tr><th>Средний</th><td>53</td><td>35</td><td>6</td><td>11</td><td>1</td><td>85.4%</td><td>97.2%</td><td>35.3%</td><td>90.9%</td></tr><tr><th>Высокий</th><td>53</td><td>36</td><td>7</td><td>10</td><td>0</td><td>83.7%</td><td>100.0%</td><td>41.2%</td><td>91.1%</td></tr><tr><th>Корпоративный</th><td>53</td><td>36</td><td>7</td><td>10</td><td>0</td><td>83.7%</td><td>100.0%</td><td>41.2%</td><td>91.1%</td></tr></table></div><h3>violence</h3><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>53</td><td>23</td><td>7</td><td>10</td><td>13</td><td>76.7%</td><td>63.9%</td><td>41.2%</td><td>69.7%</td></tr><tr><th>Средний</th><td>53</td><td>32</td><td>11</td><td>6</td><td>4</td><td>74.4%</td><td>88.9%</td><td>64.7%</td><td>81.0%</td></tr><tr><th>Высокий</th><td>53</td><td>32</td><td>12</td><td>5</td><td>4</td><td>72.7%</td><td>88.9%</td><td>70.6%</td><td>80.0%</td></tr><tr><th>Корпоративный</th><td>53</td><td>32</td><td>17</td><td>0</td><td>4</td><td>65.3%</td><td>88.9%</td><td>100.0%</td><td>75.3%</td></tr></table></div><h2>Внешний тест · восемь категорий</h2><p>Неизвестные метки исключены. Ста��дартные пороги категорий также подобраны только на внутренней части.</p><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>sexual</th><td>984</td><td>228</td><td>77</td><td>670</td><td>9</td><td>74.8%</td><td>96.2%</td><td>10.3%</td><td>84.1%</td></tr><tr><th>hate</th><td>771</td><td>158</td><td>196</td><td>413</td><td>4</td><td>44.6%</td><td>97.5%</td><td>32.2%</td><td>61.2%</td></tr><tr><th>violence</th><td>1450</td><td>86</td><td>189</td><td>1167</td><td>8</td><td>31.3%</td><td>91.5%</td><td>13.9%</td><td>46.6%</td></tr><tr><th>harassment</th><td>1444</td><td>67</td><td>443</td><td>925</td><td>9</td><td>13.1%</td><td>88.2%</td><td>32.4%</td><td>22.9%</td></tr><tr><th>self_harm</th><td>1447</td><td>43</td><td>106</td><td>1290</td><td>8</td><td>28.9%</td><td>84.3%</td><td>7.6%</td><td>43.0%</td></tr><tr><th>sexual_minors</th><td>994</td><td>85</td><td>212</td><td>697</td><td>0</td><td>28.6%</td><td>100.0%</td><td>23.3%</td><td>44.5%</td></tr><tr><th>hate_threatening</th><td>761</td><td>36</td><td>69</td><td>651</td><td>5</td><td>34.3%</td><td>87.8%</td><td>9.6%</td><td>49.3%</td></tr><tr><th>violence_graphic</th><td>1447</td><td>21</td><td>147</td><td>1276</td><td>3</td><td>12.5%</td><td>87.5%</td><td>10.3%</td><td>21.9%</td></tr></table></div><p>Лимит модели: 512 токенов. Обрезано 76 из 1680 внешних текстов. Они остаются в основном тесте, отражая реальный режим работы.</p><h2>Внешний тест · проекция четырёх уровней</h2><p>Внешний набор не размечен по нашим четырём политикам. Здесь сравнивается объединение доступных бинарных категорий. Правило мата и три отсутствующие категории исключены; это не полная точность уровня модерации.</p><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>800</td><td>202</td><td>210</td><td>323</td><td>65</td><td>49.0%</td><td>75.7%</td><td>39.4%</td><td>59.5%</td></tr><tr><th>Средний</th><td>859</td><td>493</td><td>196</td><td>141</td><td>29</td><td>71.6%</td><td>94.4%</td><td>58.2%</td><td>81.4%</td></tr><tr><th>Высокий</th><td>859</td><td>504</td><td>220</td><td>117</td><td>18</td><td>69.6%</td><td>96.6%</td><td>65.3%</td><td>80.9%</td></tr><tr><th>Корпоративный</th><td>859</td><td>505</td><td>222</td><td>115</td><td>17</td><td>69.5%</td><td>96.7%</td><td>65.9%</td><td>80.9%</td></tr></table></div><h2>Внутренний тест по языкам</h2><details><summary>be</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>24</td><td>6</td><td>2</td><td>15</td><td>1</td><td>75.0%</td><td>85.7%</td><td>11.8%</td><td>80.0%</td></tr><tr><th>Средний</th><td>24</td><td>9</td><td>5</td><td>7</td><td>3</td><td>64.3%</td><td>75.0%</td><td>41.7%</td><td>69.2%</td></tr><tr><th>Высокий</th><td>24</td><td>14</td><td>3</td><td>6</td><td>1</td><td>82.4%</td><td>93.3%</td><td>33.3%</td><td>87.5%</td></tr><tr><th>Корпоративный</th><td>24</td><td>15</td><td>2</td><td>6</td><td>1</td><td>88.2%</td><td>93.8%</td><td>25.0%</td><td>90.9%</td></tr></table></div></details><details><summary>cs</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>23</td><td>7</td><td>0</td><td>16</td><td>0</td><td>100.0%</td><td>100.0%</td><td>0.0%</td><td>100.0%</td></tr><tr><th>Средний</th><td>23</td><td>11</td><td>3</td><td>8</td><td>1</td><td>78.6%</td><td>91.7%</td><td>27.3%</td><td>84.6%</td></tr><tr><th>Высокий</th><td>23</td><td>15</td><td>2</td><td>6</td><td>0</td><td>88.2%</td><td>100.0%</td><td>25.0%</td><td>93.8%</td></tr><tr><th>Корпоративный</th><td>23</td><td>15</td><td>2</td><td>6</td><td>0</td><td>88.2%</td><td>100.0%</td><td>25.0%</td><td>93.8%</td></tr></table></div></details><details><summary>de</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>24</td><td>7</td><td>1</td><td>14</td><td>2</td><td>87.5%</td><td>77.8%</td><td>6.7%</td><td>82.4%</td></tr><tr><th>Средний</th><td>24</td><td>13</td><td>1</td><td>9</td><td>1</td><td>92.9%</td><td>92.9%</td><td>10.0%</td><td>92.9%</td></tr><tr><th>Высокий</th><td>24</td><td>15</td><td>0</td><td>8</td><td>1</td><td>100.0%</td><td>93.8%</td><td>0.0%</td><td>96.8%</td></tr><tr><th>Корпоративный</th><td>24</td><td>15</td><td>0</td><td>8</td><td>1</td><td>100.0%</td><td>93.8%</td><td>0.0%</td><td>96.8%</td></tr></table></div></details><details><summary>en</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>25</td><td>8</td><td>2</td><td>15</td><td>0</td><td>80.0%</td><td>100.0%</td><td>11.8%</td><td>88.9%</td></tr><tr><th>Средний</th><td>25</td><td>12</td><td>0</td><td>13</td><td>0</td><td>100.0%</td><td>100.0%</td><td>0.0%</td><td>100.0%</td></tr><tr><th>Высокий</th><td>25</td><td>14</td><td>0</td><td>11</td><td>0</td><td>100.0%</td><td>100.0%</td><td>0.0%</td><td>100.0%</td></tr><tr><th>Корпоративный</th><td>25</td><td>14</td><td>0</td><td>10</td><td>1</td><td>100.0%</td><td>93.3%</td><td>0.0%</td><td>96.6%</td></tr></table></div></details><details><summary>es</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>25</td><td>6</td><td>2</td><td>15</td><td>2</td><td>75.0%</td><td>75.0%</td><td>11.8%</td><td>75.0%</td></tr><tr><th>Средний</th><td>25</td><td>13</td><td>2</td><td>9</td><td>1</td><td>86.7%</td><td>92.9%</td><td>18.2%</td><td>89.7%</td></tr><tr><th>Высокий</th><td>25</td><td>16</td><td>1</td><td>6</td><td>2</td><td>94.1%</td><td>88.9%</td><td>14.3%</td><td>91.4%</td></tr><tr><th>Корпоративный</th><td>25</td><td>17</td><td>0</td><td>6</td><td>2</td><td>100.0%</td><td>89.5%</td><td>0.0%</td><td>94.4%</td></tr></table></div></details><details><summary>fi</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>24</td><td>5</td><td>3</td><td>13</td><td>3</td><td>62.5%</td><td>62.5%</td><td>18.8%</td><td>62.5%</td></tr><tr><th>Средний</th><td>24</td><td>11</td><td>4</td><td>8</td><td>1</td><td>73.3%</td><td>91.7%</td><td>33.3%</td><td>81.5%</td></tr><tr><th>Высокий</th><td>24</td><td>13</td><td>3</td><td>6</td><td>2</td><td>81.2%</td><td>86.7%</td><td>33.3%</td><td>83.9%</td></tr><tr><th>Корпоративный</th><td>24</td><td>13</td><td>3</td><td>6</td><td>2</td><td>81.2%</td><td>86.7%</td><td>33.3%</td><td>83.9%</td></tr></table></div></details><details><summary>fr</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>23</td><td>7</td><td>3</td><td>12</td><td>1</td><td>70.0%</td><td>87.5%</td><td>20.0%</td><td>77.8%</td></tr><tr><th>Средний</th><td>23</td><td>12</td><td>4</td><td>7</td><td>0</td><td>75.0%</td><td>100.0%</td><td>36.4%</td><td>85.7%</td></tr><tr><th>Высокий</th><td>23</td><td>16</td><td>0</td><td>7</td><td>0</td><td>100.0%</td><td>100.0%</td><td>0.0%</td><td>100.0%</td></tr><tr><th>Корпоративный</th><td>23</td><td>16</td><td>1</td><td>6</td><td>0</td><td>94.1%</td><td>100.0%</td><td>14.3%</td><td>97.0%</td></tr></table></div></details><details><summary>it</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>25</td><td>7</td><td>4</td><td>13</td><td>1</td><td>63.6%</td><td>87.5%</td><td>23.5%</td><td>73.7%</td></tr><tr><th>Средний</th><td>25</td><td>13</td><td>4</td><td>8</td><td>0</td><td>76.5%</td><td>100.0%</td><td>33.3%</td><td>86.7%</td></tr><tr><th>Высокий</th><td>25</td><td>16</td><td>2</td><td>7</td><td>0</td><td>88.9%</td><td>100.0%</td><td>22.2%</td><td>94.1%</td></tr><tr><th>Корпоративный</th><td>25</td><td>16</td><td>2</td><td>7</td><td>0</td><td>88.9%</td><td>100.0%</td><td>22.2%</td><td>94.1%</td></tr></table></div></details><details><summary>ja</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>24</td><td>6</td><td>1</td><td>14</td><td>3</td><td>85.7%</td><td>66.7%</td><td>6.7%</td><td>75.0%</td></tr><tr><th>Средний</th><td>24</td><td>8</td><td>4</td><td>8</td><td>4</td><td>66.7%</td><td>66.7%</td><td>33.3%</td><td>66.7%</td></tr><tr><th>Высокий</th><td>24</td><td>12</td><td>3</td><td>7</td><td>2</td><td>80.0%</td><td>85.7%</td><td>30.0%</td><td>82.8%</td></tr><tr><th>Корпоративный</th><td>24</td><td>13</td><td>2</td><td>7</td><td>2</td><td>86.7%</td><td>86.7%</td><td>22.2%</td><td>86.7%</td></tr></table></div></details><details><summary>kk</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>25</td><td>6</td><td>3</td><td>13</td><td>3</td><td>66.7%</td><td>66.7%</td><td>18.8%</td><td>66.7%</td></tr><tr><th>Средний</th><td>25</td><td>12</td><td>5</td><td>7</td><td>1</td><td>70.6%</td><td>92.3%</td><td>41.7%</td><td>80.0%</td></tr><tr><th>Высокий</th><td>25</td><td>17</td><td>2</td><td>6</td><td>0</td><td>89.5%</td><td>100.0%</td><td>25.0%</td><td>94.4%</td></tr><tr><th>Корпоративный</th><td>25</td><td>17</td><td>3</td><td>5</td><td>0</td><td>85.0%</td><td>100.0%</td><td>37.5%</td><td>91.9%</td></tr></table></div></details><details><summary>ko</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>24</td><td>6</td><td>2</td><td>15</td><td>1</td><td>75.0%</td><td>85.7%</td><td>11.8%</td><td>80.0%</td></tr><tr><th>Средний</th><td>24</td><td>10</td><td>3</td><td>10</td><td>1</td><td>76.9%</td><td>90.9%</td><td>23.1%</td><td>83.3%</td></tr><tr><th>Высокий</th><td>24</td><td>14</td><td>2</td><td>7</td><td>1</td><td>87.5%</td><td>93.3%</td><td>22.2%</td><td>90.3%</td></tr><tr><th>Корпоративный</th><td>24</td><td>15</td><td>1</td><td>7</td><td>1</td><td>93.8%</td><td>93.8%</td><td>12.5%</td><td>93.8%</td></tr></table></div></details><details><summary>lv</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>24</td><td>5</td><td>1</td><td>15</td><td>3</td><td>83.3%</td><td>62.5%</td><td>6.2%</td><td>71.4%</td></tr><tr><th>Средний</th><td>24</td><td>9</td><td>1</td><td>11</td><td>3</td><td>90.0%</td><td>75.0%</td><td>8.3%</td><td>81.8%</td></tr><tr><th>Высокий</th><td>24</td><td>13</td><td>0</td><td>8</td><td>3</td><td>100.0%</td><td>81.2%</td><td>0.0%</td><td>89.7%</td></tr><tr><th>Корпоративный</th><td>24</td><td>13</td><td>0</td><td>8</td><td>3</td><td>100.0%</td><td>81.2%</td><td>0.0%</td><td>89.7%</td></tr></table></div></details><details><summary>pl</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>23</td><td>5</td><td>4</td><td>12</td><td>2</td><td>55.6%</td><td>71.4%</td><td>25.0%</td><td>62.5%</td></tr><tr><th>Средний</th><td>23</td><td>11</td><td>3</td><td>7</td><td>2</td><td>78.6%</td><td>84.6%</td><td>30.0%</td><td>81.5%</td></tr><tr><th>Высокий</th><td>23</td><td>12</td><td>2</td><td>6</td><td>3</td><td>85.7%</td><td>80.0%</td><td>25.0%</td><td>82.8%</td></tr><tr><th>Корпоративный</th><td>23</td><td>12</td><td>2</td><td>5</td><td>4</td><td>85.7%</td><td>75.0%</td><td>28.6%</td><td>80.0%</td></tr></table></div></details><details><summary>ru</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>24</td><td>7</td><td>3</td><td>14</td><td>0</td><td>70.0%</td><td>100.0%</td><td>17.6%</td><td>82.4%</td></tr><tr><th>Средний</th><td>24</td><td>11</td><td>3</td><td>8</td><td>2</td><td>78.6%</td><td>84.6%</td><td>27.3%</td><td>81.5%</td></tr><tr><th>Высокий</th><td>24</td><td>14</td><td>0</td><td>8</td><td>2</td><td>100.0%</td><td>87.5%</td><td>0.0%</td><td>93.3%</td></tr><tr><th>Корпоративный</th><td>24</td><td>14</td><td>0</td><td>8</td><td>2</td><td>100.0%</td><td>87.5%</td><td>0.0%</td><td>93.3%</td></tr></table></div></details><details><summary>sv</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>25</td><td>7</td><td>0</td><td>17</td><td>1</td><td>100.0%</td><td>87.5%</td><td>0.0%</td><td>93.3%</td></tr><tr><th>Средний</th><td>25</td><td>12</td><td>2</td><td>9</td><td>2</td><td>85.7%</td><td>85.7%</td><td>18.2%</td><td>85.7%</td></tr><tr><th>Высокий</th><td>25</td><td>16</td><td>0</td><td>8</td><td>1</td><td>100.0%</td><td>94.1%</td><td>0.0%</td><td>97.0%</td></tr><tr><th>Корпоративный</th><td>25</td><td>17</td><td>1</td><td>7</td><td>0</td><td>94.4%</td><td>100.0%</td><td>12.5%</td><td>97.1%</td></tr></table></div></details><details><summary>uk</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>25</td><td>7</td><td>2</td><td>16</td><td>0</td><td>77.8%</td><td>100.0%</td><td>11.1%</td><td>87.5%</td></tr><tr><th>Средний</th><td>25</td><td>12</td><td>3</td><td>9</td><td>1</td><td>80.0%</td><td>92.3%</td><td>25.0%</td><td>85.7%</td></tr><tr><th>Высокий</th><td>25</td><td>14</td><td>2</td><td>7</td><td>2</td><td>87.5%</td><td>87.5%</td><td>22.2%</td><td>87.5%</td></tr><tr><th>Корпоративный</th><td>25</td><td>15</td><td>2</td><td>6</td><td>2</td><td>88.2%</td><td>88.2%</td><td>25.0%</td><td>88.2%</td></tr></table></div></details><details><summary>zh</summary><div class="scroll"><table><tr><th>Уровень / категория</th><th>n</th><th>TP</th><th>FP</th><th>TN</th><th>FN</th><th>precision</th><th>recall</th><th>false_positive_rate</th><th>f1</th></tr><tr><th>Лёгкий</th><td>25</td><td>7</td><td>2</td><td>15</td><td>1</td><td>77.8%</td><td>87.5%</td><td>11.8%</td><td>82.4%</td></tr><tr><th>Средний</th><td>25</td><td>10</td><td>4</td><td>8</td><td>3</td><td>71.4%</td><td>76.9%</td><td>33.3%</td><td>74.1%</td></tr><tr><th>Высокий</th><td>25</td><td>14</td><td>2</td><td>7</td><td>2</td><td>87.5%</td><td>87.5%</td><td>22.2%</td><td>87.5%</td></tr><tr><th>Корпоративный</th><td>25</td><td>15</td><td>1</td><td>7</td><td>2</td><td>93.8%</td><td>88.2%</td><td>12.5%</td><td>90.9%</td></tr></table></div></details><h2>Карта покрытия 11 × 17</h2><p>Положительные примеры по исходной оценке ≥ 0,5. В ячейках: калибровка / отложенная проверка. Переводы связаны общими группами.</p><div class="scroll"><table><tr><th>Категория</th><th>en</th><th>de</th><th>fr</th><th>es</th><th>it</th><th>sv</th><th>fi</th><th>pl</th><th>cs</th><th>lv</th><th>zh</th><th>ja</th><th>ko</th><th>ru</th><th>uk</th><th>be</th><th>kk</th></tr><tr><th>harassment</th><td>10 / 5</td><td>17 / 6</td><td>12 / 9</td><td>17 / 8</td><td>12 / 7</td><td>14 / 9</td><td>11 / 7</td><td>14 / 7</td><td>16 / 6</td><td>17 / 7</td><td>15 / 6</td><td>13 / 6</td><td>12 / 7</td><td>17 / 6</td><td>14 / 7</td><td>14 / 6</td><td>12 / 9</td></tr><tr><th>harassment_threatening</th><td>3 / 4</td><td>5 / 4</td><td>4 / 3</td><td>3 / 4</td><td>4 / 3</td><td>4 / 3</td><td>4 / 3</td><td>4 / 3</td><td>5 / 3</td><td>4 / 3</td><td>4 / 3</td><td>3 / 5</td><td>2 / 4</td><td>3 / 3</td><td>3 / 3</td><td>3 / 4</td><td>3 / 4</td></tr><tr><th>hate</th><td>3 / 4</td><td>4 / 3</td><td>3 / 4</td><td>4 / 3</td><td>4 / 3</td><td>3 / 4</td><td>4 / 3</td><td>3 / 4</td><td>7 / 3</td><td>4 / 4</td><td>3 / 4</td><td>5 / 4</td><td>3 / 3</td><td>4 / 3</td><td>3 / 3</td><td>5 / 3</td><td>2 / 5</td></tr><tr><th>hate_threatening</th><td>1 / 3</td><td>3 / 2</td><td>2 / 2</td><td>2 / 2</td><td>2 / 2</td><td>2 / 2</td><td>2 / 2</td><td>1 / 2</td><td>2 / 2</td><td>2 / 2</td><td>2 / 2</td><td>2 / 2</td><td>1 / 2</td><td>1 / 3</td><td>1 / 2</td><td>2 / 2</td><td>1 / 3</td></tr><tr><th>self_harm</th><td>3 / 3</td><td>2 / 4</td><td>4 / 2</td><td>5 / 2</td><td>4 / 3</td><td>4 / 2</td><td>3 / 4</td><td>4 / 2</td><td>5 / 3</td><td>3 / 2</td><td>5 / 2</td><td>4 / 3</td><td>4 / 2</td><td>3 / 3</td><td>4 / 3</td><td>4 / 2</td><td>3 / 3</td></tr><tr><th>self_harm_instructions</th><td>2 / 1</td><td>1 / 2</td><td>3 / 1</td><td>3 / 1</td><td>2 / 2</td><td>3 / 1</td><td>2 / 2</td><td>2 / 1</td><td>2 / 2</td><td>1 / 2</td><td>3 / 1</td><td>2 / 2</td><td>2 / 1</td><td>2 / 1</td><td>1 / 2</td><td>2 / 1</td><td>1 / 2</td></tr><tr><th>self_harm_intent</th><td>3 / 3</td><td>2 / 4</td><td>3 / 2</td><td>5 / 2</td><td>4 / 3</td><td>4 / 2</td><td>3 / 4</td><td>4 / 2</td><td>4 / 2</td><td>3 / 2</td><td>4 / 2</td><td>4 / 3</td><td>4 / 3</td><td>3 / 3</td><td>4 / 2</td><td>3 / 2</td><td>3 / 3</td></tr><tr><th>sexual</th><td>6 / 3</td><td>7 / 3</td><td>10 / 3</td><td>7 / 4</td><td>10 / 5</td><td>7 / 4</td><td>9 / 4</td><td>8 / 3</td><td>7 / 3</td><td>9 / 5</td><td>7 / 4</td><td>6 / 3</td><td>9 / 5</td><td>8 / 2</td><td>6 / 3</td><td>12 / 2</td><td>10 / 3</td></tr><tr><th>sexual_minors</th><td>4 / 2</td><td>5 / 1</td><td>5 / 2</td><td>5 / 2</td><td>5 / 2</td><td>4 / 2</td><td>5 / 2</td><td>5 / 1</td><td>4 / 1</td><td>4 / 2</td><td>4 / 2</td><td>4 / 1</td><td>5 / 1</td><td>5 / 1</td><td>5 / 1</td><td>5 / 1</td><td>5 / 1</td></tr><tr><th>violence</th><td>11 / 5</td><td>9 / 5</td><td>12 / 3</td><td>8 / 5</td><td>10 / 4</td><td>7 / 4</td><td>10 / 5</td><td>10 / 4</td><td>7 / 4</td><td>8 / 4</td><td>9 / 6</td><td>7 / 6</td><td>8 / 4</td><td>5 / 6</td><td>7 / 4</td><td>9 / 7</td><td>10 / 5</td></tr><tr><th>violence_graphic</th><td>5 / 1</td><td>3 / 3</td><td>5 / 1</td><td>3 / 2</td><td>4 / 1</td><td>3 / 2</td><td>4 / 1</td><td>4 / 1</td><td>3 / 2</td><td>4 / 1</td><td>3 / 2</td><td>3 / 2</td><td>4 / 1</td><td>2 / 2</td><td>3 / 2</td><td>3 / 2</td><td>5 / 1</td></tr></table></div><p>TP — верно заблокировано; FP — лишняя блокировка; TN — верно пропущено; FN — пропуск запрещённого. Precision — доля верных блокировок; recall — доля обнаруженного запрещённого.</p><p><a href="evaluation.json">Полные метрики JSON</a> · <a href="thresholds.json">Зафиксированные пороги</a> · <a href="scores_without_text.jsonl">Оценки по строкам без исходных текстов</a></p></html>

evaluation/thresholds.json

@@ -0,0 +1,78 @@
+{
+  "status": "fitted_internal_only",
+  "method": "group-weighted balanced error threshold search; nested profiles",
+  "fit_rows": 828,
+  "fit_groups": 638,
+  "thresholds": {
+    "light": {
+      "harassment": null,
+      "harassment_threatening": 0.12,
+      "hate": 0.26,
+      "hate_threatening": 0.01,
+      "self_harm": null,
+      "self_harm_instructions": 0.23,
+      "self_harm_intent": null,
+      "sexual": null,
+      "sexual_minors": 0.02,
+      "violence": null,
+      "violence_graphic": 0.02
+    },
+    "medium": {
+      "harassment": 0.45,
+      "harassment_threatening": 0.03,
+      "hate": 0.060000000000000005,
+      "hate_threatening": 0.01,
+      "self_harm": 0.12,
+      "self_harm_instructions": 0.01,
+      "self_harm_intent": 0.02,
+      "sexual": 0.59,
+      "sexual_minors": 0.01,
+      "violence": 0.09999999999999999,
+      "violence_graphic": 0.02
+    },
+    "high": {
+      "harassment": 0.24000000000000002,
+      "harassment_threatening": 0.03,
+      "hate": 0.05,
+      "hate_threatening": 0.01,
+      "self_harm": 0.03,
+      "self_harm_instructions": 0.01,
+      "self_harm_intent": 0.02,
+      "sexual": 0.34,
+      "sexual_minors": 0.01,
+      "violence": 0.09999999999999999,
+      "violence_graphic": 0.02
+    },
+    "corporate": {
+      "harassment": 0.24000000000000002,
+      "harassment_threatening": 0.01,
+      "hate": 0.05,
+      "hate_threatening": 0.01,
+      "self_harm": 0.02,
+      "self_harm_instructions": 0.01,
+      "self_harm_intent": 0.01,
+      "sexual": 0.29000000000000004,
+      "sexual_minors": 0.01,
+      "violence": 0.09,
+      "violence_graphic": 0.02
+    }
+  },
+  "standard_category_thresholds": {
+    "harassment": 0.25,
+    "harassment_threatening": 0.02,
+    "hate": 0.03,
+    "hate_threatening": 0.01,
+    "self_harm": 0.01,
+    "self_harm_instructions": 0.01,
+    "self_harm_intent": 0.01,
+    "sexual": 0.11,
+    "sexual_minors": 0.01,
+    "violence": 0.09,
+    "violence_graphic": 0.02
+  },
+  "probabilities_calibrated": false,
+  "checkpoint_sha256": "4218fb73825ff2fbbdd897a4346dd67141b15fc63d1811413929dddb88d408d3",
+  "fit_ids_sha256": "852388954817cf5ef3047eae79e8a1480cc60a1311673e439a02fbb9511cca23",
+  "policy": "Reference four-profile policy plus finite profanity rule. Research intent checks scored separately.",
+  "limitations": "Weak teacher-derived and synthetic labels. No new detector heads were trained."
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4218fb73825ff2fbbdd897a4346dd67141b15fc63d1811413929dddb88d408d3
+size 1018281676

moderation03.py

@@ -0,0 +1,82 @@
+"""Portable inference for the exact FP32 epoch-5 head used in the published evaluation."""
+import gc
+import json
+import os
+from pathlib import Path
+import torch
+from huggingface_hub import snapshot_download
+from safetensors.torch import load_file
+from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration
+from moderation_model import ModerationHead
+from moderation_policy import LANGS, PROFILES, profanity
+from categories import CATEGORIES
+
+MODEL_ID='ifmain/Moderation-03'
+
+def apply_policy(scores, thresholds, *, block_profanity=False, text='', language='en'):
+    if set(scores)!=set(CATEGORIES) or set(thresholds)!=set(CATEGORIES):
+        raise ValueError('Exactly 11 named categories are required')
+    if language not in LANGS:raise ValueError('Unsupported language')
+    for value in thresholds.values():
+        if value is not None and not 0<=value<=1:raise ValueError('Thresholds must be in [0,1] or null (disabled)')
+    tags={c:thresholds[c] is not None and scores[c]>=thresholds[c] for c in CATEGORIES}
+    reasons=[c for c,flagged in tags.items() if flagged]
+    lexical=block_profanity and profanity(text,language)
+    if lexical:reasons.append('profanity_rule')
+    return {'block':bool(reasons),'tags':tags,'reasons':reasons,'thresholds':thresholds,
+            'profanity_rule_enabled':block_profanity,'profanity_rule_triggered':bool(lexical)}
+
+class Moderation03:
+    def __init__(self, model_id=MODEL_ID, device=None, backbone_path=None):
+        self.device=device or ('cuda' if torch.cuda.is_available() else 'cpu')
+        torch.set_num_threads(int(os.getenv('TORCH_NUM_THREADS','4')))
+        torch.backends.mha.set_fastpath_enabled(False)
+        path=Path(model_id)
+        self.root=path if path.is_dir() else Path(snapshot_download(model_id,allow_patterns=[
+            'config.json','model.safetensors','calibration/thresholds.json']))
+        self.config=json.loads((self.root/'config.json').read_text(encoding='utf-8'))
+        self.calibration=json.loads((self.root/'calibration/thresholds.json').read_text(encoding='utf-8'))
+        backbone=backbone_path or self.config['backbone']
+        local=Path(backbone).is_dir()
+        kwargs={'local_files_only':True} if local else {'revision':self.config['backbone_revision']}
+        self.tokenizer=AutoTokenizer.from_pretrained(backbone,**kwargs)
+        self.tokenizer.padding_side='right'
+        if self.tokenizer.pad_token_id is None:self.tokenizer.pad_token=self.tokenizer.eos_token
+        full=Qwen3_5ForConditionalGeneration.from_pretrained(backbone,dtype=torch.bfloat16,
+                attn_implementation='sdpa',**kwargs)
+        self.backbone=full.model.language_model
+        del full;gc.collect()
+        self.backbone.requires_grad_(False).eval().to(self.device)
+        self.head=ModerationHead(self.config)
+        self.head.load_state_dict(load_file(str(self.root/'model.safetensors'),device='cpu'))
+        self.head.requires_grad_(False).eval().to(self.device)
+
+    @torch.inference_mode()
+    def scores(self,text):
+        if not isinstance(text,str) or not text.strip():raise ValueError('Enter non-empty text')
+        if len(text)>50000:raise ValueError('Maximum input length is 50,000 characters')
+        ids=self.tokenizer(text,add_special_tokens=False,truncation=False)['input_ids']
+        total=len(ids);ids=ids[:self.config['max_length']] or [self.tokenizer.eos_token_id]
+        encoded={'input_ids':torch.tensor([ids],device=self.device),
+                 'attention_mask':torch.ones((1,len(ids)),dtype=torch.long,device=self.device)}
+        hidden=self.backbone(**encoded,use_cache=False,return_dict=True).last_hidden_state
+        with torch.autocast(self.device,dtype=torch.bfloat16):
+            logits=self.head(hidden.float(),encoded['attention_mask'])
+        if not bool(torch.isfinite(logits).all()):raise ValueError('Non-finite scores')
+        values=logits.float().sigmoid()[0].cpu().tolist()
+        return {'raw_scores':dict(zip(CATEGORIES,values)), 'input_tokens':total,
+                'processed_tokens':len(ids),'truncated':total>self.config['max_length'],
+                'checkpoint_epoch':5,'score_type':'uncalibrated sigmoid score'}
+
+    def predict(self,text,language='en',preset=None,thresholds=None,block_profanity=False):
+        if language not in LANGS:raise ValueError('Unsupported language')
+        if preset is not None and preset not in PROFILES:raise ValueError('Unknown preset')
+        if preset is not None and thresholds is not None:raise ValueError('Choose preset or custom thresholds')
+        result=self.scores(text)
+        if preset is not None:
+            thresholds=self.calibration['thresholds'][preset]
+            block_profanity=preset in ['high','corporate']
+        result['preset']=preset
+        result['policy']=None if thresholds is None else apply_policy(result['raw_scores'],thresholds,
+                block_profanity=block_profanity,text=text,language=language)
+        return result

moderation_model.py

@@ -0,0 +1,70 @@
+"""Trainable transformer head over the full sequence of frozen Qwen hidden states."""
+import json
+from pathlib import Path
+import torch
+from torch import nn
+from torch.utils.checkpoint import checkpoint
+
+DEFAULT_CONFIG = Path(__file__).resolve().parent / 'models/moderation_head/config.json'
+
+
+def load_config(path=DEFAULT_CONFIG):
+    return json.loads(Path(path).read_text(encoding='utf-8'))
+
+
+class ModerationHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.input_norm = nn.LayerNorm(config['input_size'])
+        self.projection = nn.Linear(config['input_size'], config['hidden_size'])
+        self.positions = nn.Embedding(config['max_length'], config['hidden_size'])
+        # Construct independently: every layer receives its own random initialization.
+        self.layers = nn.ModuleList([
+            nn.TransformerEncoderLayer(
+                d_model=config['hidden_size'], nhead=config['num_heads'],
+                dim_feedforward=config['intermediate_size'], dropout=config['dropout'],
+                activation='gelu', batch_first=True, norm_first=True)
+            for _ in range(config['num_layers'])
+        ])
+        self.final_norm = nn.LayerNorm(config['hidden_size'])
+        self.classifier = nn.Linear(config['hidden_size'], len(config['categories']))
+
+    def forward(self, hidden_states, attention_mask):
+        if hidden_states.ndim != 3 or hidden_states.shape[-1] != self.config['input_size']:
+            raise ValueError('Expected [batch, sequence, Qwen_hidden_size]')
+        if hidden_states.shape[:2] != attention_mask.shape:
+            raise ValueError('Mask shape mismatch')
+        if hidden_states.shape[1] > self.config['max_length']:
+            raise ValueError('Sequence exceeds configured maximum')
+        mask = attention_mask.bool()
+        if not bool(mask.any(dim=1).all()):
+            raise ValueError('Every example requires at least one non-padding token')
+        x = self.projection(self.input_norm(hidden_states))
+        positions = (mask.long().cumsum(dim=1) - 1).clamp_min(0)
+        x = x + self.positions(positions)
+        padding = ~mask
+        for layer in self.layers:
+            if self.training and self.config.get('gradient_checkpointing', False):
+                x = checkpoint(layer, x, src_key_padding_mask=padding, use_reentrant=False)
+            else:
+                x = layer(x, src_key_padding_mask=padding)
+        x = self.final_norm(x)
+        pooled = (x * mask.unsqueeze(-1)).sum(dim=1) / mask.sum(dim=1, keepdim=True)
+        return self.classifier(pooled)
+
+    @torch.inference_mode()
+    def predict_scores(self, hidden_states, attention_mask):
+        if self.training:
+            raise RuntimeError('Call eval() before inference')
+        probabilities = self(hidden_states, attention_mask).float().sigmoid()
+        return [dict(zip(self.config['categories'], row)) for row in probabilities.cpu().tolist()]
+
+
+def apply_profile(scores, thresholds):
+    """Future calibrated thresholds must be provided explicitly for all categories."""
+    if thresholds is None or set(scores) != set(thresholds):
+        raise ValueError('This profile has not been calibrated for all 11 categories')
+    if any(not 0 <= v <= 1 for v in thresholds.values()):
+        raise ValueError('Thresholds must lie in [0, 1]')
+    return {name: score >= thresholds[name] for name, score in scores.items()}

moderation_policy.py

@@ -0,0 +1,60 @@
+"""Four nested policies over 11 raw outputs plus a separate explicit profanity rule."""
+import re
+import unicodedata
+from categories import CATEGORIES
+
+LANGS='en de fr es it sv fi pl cs lv zh ja ko ru uk be kk'.split()
+PROFILES=['light','medium','high','corporate']
+# These are policy definitions, not measured model probability thresholds.
+REFERENCE={
+ 'light':[None,.60,.80,.50,None,.60,None,None,.50,None,.60],
+ 'medium':[.55,.45,.60,.40,.85,.45,.85,.80,.40,.85,.45],
+ 'high':[.30,.30,.40,.25,.65,.30,.65,.50,.25,.60,.30],
+ 'corporate':[.15,.20,.20,.15,.45,.20,.45,.20,.15,.30,.20],
+}
+TERMS={
+ 'en':['fuck','fucking','shit','bullshit','motherfucker'],
+ 'de':['scheiße','scheisse','verdammt','arschloch'],
+ 'fr':['putain','merde','bordel'], 'es':['mierda','joder','coño'],
+ 'it':['cazzo','merda','vaffanculo'], 'sv':['jävla','jävlar','helvete'],
+ 'fi':['vittu','perkele','saatana'], 'pl':['kurwa','kurwy','pierdolić'],
+ 'cs':['kurva','kurvy','do prdele'], 'lv':['bļeģ','bļadj','pizģec'],
+ 'zh':['他妈的','操你妈','傻逼'], 'ja':['くそ','クソ','ちくしょう'],
+ 'ko':['씨발','시발','좆'], 'ru':['блядь','блять','хуй','хуя','пиздец','ебать','ебаный'],
+ 'uk':['блядь','блять','хуй','хуя','пиздець','їбаний'],
+ 'be':['блядзь','блядь','хуй','піздзец'], 'kk':['қотақ','блядь','пиздец'],
+}
+HOMO=str.maketrans({'а':'a','е':'e','ё':'e','о':'o','р':'p','с':'c','х':'x','у':'y','і':'i'})
+
+
+def norm(text):
+    return unicodedata.normalize('NFKC',text).casefold().translate(HOMO)
+
+
+def pattern(word, lang):
+    # Inserted separators and common digit substitutions are accepted inside known words.
+    variants={'a':'a4','e':'e3','i':'i1','o':'o0','s':'s5','t':'t7','b':'b8','g':'g9','z':'z2'}
+    letters=[]
+    for ch in norm(word):
+        if ch.isspace():continue
+        letters.append('['+re.escape(variants.get(ch,ch))+']')
+    core=r'[\s_.\-]*'.join(letters)
+    return re.compile(core if lang in {'zh','ja','ko'} else r'(?<!\w)'+core+r'(?!\w)')
+
+PATTERNS={lang:[pattern(word,lang) for word in terms] for lang,terms in TERMS.items()}
+
+
+def profanity(text,language):
+    normalized=norm(text)
+    langs=[language] if language in PATTERNS else LANGS
+    return any(p.search(normalized) for lang in langs for p in PATTERNS[lang])
+
+
+def decisions(scores, profane, thresholds=None):
+    thresholds=thresholds or {p:dict(zip(CATEGORIES,REFERENCE[p])) for p in PROFILES}
+    result={}
+    for profile in PROFILES:
+        reasons=[c for c in CATEGORIES if thresholds[profile][c] is not None and scores[c]>=thresholds[profile][c]]
+        if profane and profile in {'high','corporate'}:reasons.append('profanity_rule')
+        result[profile]={'block':bool(reasons),'reasons':reasons}
+    return result

requirements.txt

@@ -0,0 +1,5 @@
+torch==2.13.0
+transformers==5.14.1
+safetensors==0.8.0
+huggingface-hub==1.26.0
+gradio==6.13.0