toxic v0.1.0

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,202 @@
+---
+license: other
+license_name: desert-ant-labs-source-available-1.0
+license_link: https://license.desertant.com/1.0
+language:
+- de
+- en
+- es
+- fr
+- it
+- nl
+- pl
+- pt
+tags:
+- hate-speech
+- content-moderation
+- text-classification
+- on-device
+- litert
+- tflite
+- core-ml
+- multilingual
+pipeline_tag: text-classification
+---
+
+# toxic: on-device multilingual hate-speech and abuse triage
+
+An on-device multilingual hate-speech / abuse triage classifier: five content
+heads (`HATEFUL`, `HARASSMENT`, `THREAT`, `SEXUAL`, `SELF_HARM`) plus
+protected-group target heads, measured on eight EU languages (en, de, fr, nl,
+pt, es, it, pl). The deployable model is **~66.7 MB** (4-bit Core ML on Apple,
+2.8 ms on an M3 Ultra Neural Engine) or **~75.5 MB** (int4 LiteRT `.tflite` on
+Android, Linux and the web). Inference is 100% on-device: no network call,
+ever.
+
+**Triage, not verdict.** Outputs are escalation signals for human review or a
+heavier local tier, not autonomous removal decisions. `SELF_HARM` and `THREAT`
+are high-precision heads intended for resource routing with a human in the
+loop. The failure modes below are published on purpose: read them before
+wiring the model into any enforcement path.
+
+## Files
+
+| File | Format | Size | Contents |
+|---|---|---:|---|
+| `toxic.tflite` | LiteRT / TFLite (int4 blockwise-32) | ~75.5 MB | Dynamic-range int4 (block-32 groups) + int8, fixed `[1, 128]` window; runs on Android, Linux and the web |
+| `toxic.mlmodelc` | Compiled Core ML (4-bit) | ~66.7 MB | Palettized (grouped-channel encoder + per-tensor embedding), fp16 compute, ANE-resident; ready to load on Apple platforms |
+| `toxic.mlpackage` | Core ML package (4-bit) | ~66.7 MB | Same model before `coremlcompiler`; recompile or inspect |
+| `toxic.pt` | PyTorch state dict | ~531 MB | Full-precision reference weights (encoder + both heads) |
+| `config.json` | JSON | tiny | Encoder + head config (base, labels, `max_len`, vocab size) |
+| `tokenizer.json`, `tokenizer_config.json` | JSON | ~2.4 MB | EU-trimmed (60,728-piece) SentencePiece-Unigram tokenizer (XLM-R lineage) |
+| `toxic_tokenizer.bin` | Binary | ~0.7 MB | Compact `RDTK` vocab (scores + pieces) for the portable SDK tokenizers |
+| `labels.json` | JSON | tiny | `id2label` / `label2id` for both heads |
+| `toxic_meta.json` | JSON | tiny | Version, languages, labels, default threshold and the per-language threshold menu |
+
+## Taxonomy
+
+Two multi-label heads over one shared encoder. Text can fire several labels at
+once; thresholds apply per label.
+
+- **Content heads (5):** `HATEFUL` (scoped to the EU notion of illegal hate
+  speech: public incitement to violence or hatred against a protected group,
+  Framework Decision 2008/913/JHA), `HARASSMENT` (targeted insult / bullying
+  of an individual), `THREAT` (threat of violence), `SEXUAL` (sexual
+  harassment / unsolicited sexual aggression), `SELF_HARM` (self-harm /
+  suicide content, for resource routing).
+- **Target heads (10):** the 2008/913/JHA protected grounds `RACE`, `COLOUR`,
+  `RELIGION`, `DESCENT`, `NATIONAL_ETHNIC_ORIGIN`, plus the extended grounds
+  `SEXUAL_ORIENTATION`, `GENDER`, `DISABILITY`, `AGE`, `OTHER`. Only
+  meaningful when `HATEFUL` fires; used for per-ground fairness reporting.
+
+## Architecture
+
+- **Encoder:** XLM-R-base lineage (12 layers, hidden 768), vocabulary trimmed
+  from 250k to 60,728 pieces (~133M params), fine-tuned with focal BCE on the
+  two heads.
+- **Static shapes:** fixed `[1, 128]` input, no dynamic control flow;
+  ANE-native reformulation for the Core ML export, dynamic-range quantization
+  for LiteRT.
+- **Runtime:** tokenize off-graph (SentencePiece-Unigram, `tokenizer.json` or
+  `toxic_tokenizer.bin`), run the encoder, apply sigmoid per label, threshold
+  per the menu below. Default threshold **0.40**.
+
+## Measured quality (the shipped artifacts, not the checkpoint)
+
+Every number below is the full Multilingual HateCheck scorer re-run on the
+exact artifact in this repo, threshold 0.40. HateCheck is eval-only: it is
+never trained on. Torch reference: 8-language mean macro-F1 **0.789**.
+
+| platform | artifact | size | 8-lang macro-F1 | delta vs torch | latency (measured) |
+| --- | --- | --: | --: | --: | --- |
+| iOS / macOS (ANE) | `toxic.mlmodelc` | 66.7 MB | 0.776 | -0.013 | 2.82 ms median (M3 Ultra ANE); 12.6 ms CPU-only |
+| Android / web | `toxic.tflite` | 75.5 MB | 0.786 | -0.003 | 8.1 ms median (8-thread desktop CPU, XNNPACK) |
+| reference | `toxic.pt` | 531 MB | 0.789 | - | n/a |
+
+Phone-class latency (iPhone 13/15, mid-range Android) is not yet measured;
+the M3 Ultra figure bounds the graph cost.
+
+Per language (threshold 0.40):
+
+| lang | torch | `toxic.tflite` (int4-b32) | `toxic.mlmodelc` (int4) |
+| --- | --: | --: | --: |
+| en | 0.810 | 0.787 | 0.771 |
+| de | 0.803 | 0.790 | 0.791 |
+| fr | 0.824 | 0.816 | 0.798 |
+| nl | 0.764 | 0.769 | 0.754 |
+| pt | 0.791 | 0.800 | 0.785 |
+| es | 0.796 | 0.799 | 0.786 |
+| it | 0.775 | 0.776 | 0.771 |
+| pl | 0.751 | 0.750 | 0.754 |
+| mean | 0.789 | 0.786 | 0.776 |
+
+## Threshold menu (per-language operating points)
+
+Measured on the torch reference (the vocabulary trim is quality-neutral).
+`macro-F1` is the HateCheck 2-class macro; `FPR` is the false-positive rate on
+non-hate cases, the number to watch if you fear over-blocking. Full sweep
+tables (0.30 to 0.70) live in the training repo (`reports/RELEASE_sweep.md`).
+
+| lang | tuned threshold | macro-F1 @ tuned | FPR @ tuned | macro-F1 @ 0.40 | FPR @ 0.40 |
+| --- | --: | --: | --: | --: | --: |
+| en | 0.40 | 0.810 | 0.281 | 0.810 | 0.281 |
+| de | 0.55 | 0.822 | 0.254 | 0.803 | 0.359 |
+| fr | 0.45 | 0.828 | 0.247 | 0.824 | 0.272 |
+| nl | 0.55 | 0.768 | 0.238 | 0.764 | 0.298 |
+| pt | 0.60 | 0.800 | 0.203 | 0.791 | 0.288 |
+| es | 0.55 | 0.803 | 0.207 | 0.796 | 0.272 |
+| it | 0.60 | 0.779 | 0.234 | 0.775 | 0.309 |
+| pl | 0.30 | 0.759 | 0.305 | 0.751 | 0.271 |
+
+Read: raising the threshold trades hate recall for a lower non-hate FPR. At
+per-language tuned thresholds the 7-EU-language average is 0.794 and four
+languages (de, fr, pt, es) clear 0.80. The same menu ships in
+`toxic_meta.json`.
+
+## Measured competitors (same scorer, same suite, shipping quants)
+
+Every runnable competitor was re-run through the identical HateCheck harness.
+Generative guards are scored on their parsed safety verdict (Unsafe = hate);
+their sizes are published Q4_K_M GGUF bytes, and their int4 quality is
+measured (bitsandbytes nf4). 7-EU = mean macro-F1 over de, fr, nl, pt, es,
+it, pl.
+
+| system | on-disk size | 7-EU avg | en | notes |
+| --- | --: | --: | --: | --- |
+| toxic `toxic.tflite` (int4-b32) | 75.5 MB | 0.786 | 0.787 | this repo, measured artifact |
+| toxic `toxic.mlmodelc` (int4) | 66.7 MB | 0.777 | 0.771 | this repo, measured artifact |
+| Llama-Guard-3-1B (int4) | 955.4 MB | 0.633 | 0.794 | verdict parsing; bf16 en 0.814 |
+| Qwen3Guard-Gen-0.6B (int4) | 484.2 MB | 0.622 | 0.750 | verdict parsing; bf16 en 0.770 |
+| Detoxify multilingual (fp32, 278M) | ~280 MB | 0.487 | 0.642 | fixed th 0.5; collapses on de/nl/pl; int8 unmeasured |
+
+toxic leads every EU language at a fraction of the size; Llama-Guard-3-1B
+still leads on English (0.794 vs 0.787 at its 12.7x size).
+
+## Failure modes (read before deploying)
+
+Publishing per-language, per-functionality error rates is part of the
+product. Weakest HateCheck cells per language (accuracy on that cell,
+flagship int8 artifact, threshold 0.40; `_nh` cells are non-hate whose error
+is a false positive):
+
+| lang | weakest cells (acc) |
+| --- | --- |
+| en | counter_ref_nh 0.51, negate_neg_nh 0.53, slur_reclaimed_nh 0.56, counter_quote_nh 0.63, slur_homonym_nh 0.63 |
+| de | negate_neg_nh 0.39, ident_pos_nh 0.45, counter_ref_nh 0.48, slur_h 0.58 |
+| fr | ident_neutral_nh 0.59, target_group_nh 0.63, ident_pos_nh 0.67, derog_impl_h 0.69 |
+| nl | slur_h 0.41, ident_pos_nh 0.47, negate_neg_nh 0.52, phrase_question_h 0.58 |
+| pt | ident_neutral_nh 0.48, target_group_nh 0.57, ident_pos_nh 0.62, slur_h 0.67 |
+| es | ident_neutral_nh 0.52, target_group_nh 0.57, ident_pos_nh 0.66, target_obj_nh 0.68 |
+| it | ident_pos_nh 0.52, ident_neutral_nh 0.53, target_group_nh 0.55, phrase_question_h 0.57 |
+| pl | ident_pos_nh 0.41, slur_h 0.56, phrase_question_h 0.66, negate_neg_nh 0.67 |
+
+The pattern: counter-speech, negated hate, reclaimed slurs and positive /
+neutral identity mentions are the dominant false-positive sources in every
+language; lexical slurs (`slur_h`) are the dominant miss in nl, de and pl.
+If your application is FPR-sensitive, use the higher per-language thresholds
+above and route positives to review instead of auto-actioning.
+
+Quantization costs, measured on the shipped artifacts:
+
+- **Core ML (iOS): palettization costs about 3 to 4 points on English.**
+  fp16 on the ANE is quality-neutral (en 0.811 vs torch 0.810); the shipped
+  4-bit palettized model lands en 0.771 and 8-language 0.776 (-1.3pp mean vs
+  torch). English carries most of the palettization loss.
+- **TFLite (Android / web): English pays -2.1pp for int4.** `toxic.tflite`
+  (int4 blockwise-32) scores en 0.787 vs 0.808 for the int8 build of the
+  same model; the 7-EU average is identical for both (0.786). If English
+  dominates your traffic and 59 MB matters less than 2pp, ask for the int8
+  build.
+- Non-hate FPR at the 0.40 default runs 0.27 to 0.36 depending on language
+  (see the threshold menu for the FPR you actually opt into).
+
+## License
+
+[Desert Ant Labs Source-Available License](https://license.desertant.com/1.0).
+Free for most apps; a commercial license is required at scale. Full terms at
+the link. Licensing: <licensing@desertant.com>.
+
+Built exclusively from commercially clean components (CC0 / CC-BY / MIT /
+Apache-2.0 training data, MIT base encoder). Attributions, dataset citations
+(GAHD, Measuring Hate Speech, GermEval 2018, MLMA and others) and generator
+credits are in `THIRD_PARTY_NOTICES.md`.

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,62 @@
+# Third-party notices: toxic
+
+`toxic` is built for commercial use. It derives from the components below,
+each used under a license that permits commercial use and derivative works.
+Their license terms apply to those components; the attributions here satisfy
+them. Evaluation suites are never trained on.
+
+## Base encoder
+
+### XLM-RoBERTa base (FacebookAI)
+- [`FacebookAI/xlm-roberta-base`](https://huggingface.co/FacebookAI/xlm-roberta-base): **MIT**.
+- Used as the encoder: vocabulary trimmed from 250k to 60,728 pieces,
+  fine-tuned for multi-label hate-speech / abuse classification.
+
+## Training datasets
+
+- [`google/civil_comments`](https://huggingface.co/datasets/google/civil_comments)
+  (Civil Comments / Jigsaw Unintended Bias): **CC0-1.0** (public domain).
+  Courtesy citation: Borkan, Dixon, Sorensen, Thain, Vasserman, "Nuanced
+  Metrics for Measuring Unintended Bias with Real Data for Text
+  Classification", WWW 2019.
+- [`ucberkeley-dlab/measuring-hate-speech`](https://huggingface.co/datasets/ucberkeley-dlab/measuring-hate-speech)
+  (Measuring Hate Speech, MHS): **CC-BY-4.0**. Attribution: D-Lab, UC
+  Berkeley. Citation: Kennedy, Bacon, Sahn, von Vacano, "Constructing
+  interval variables via faceted Rasch measurement and multitask deep
+  learning: a hate speech application", 2020.
+- [GAHD (German Adversarial Hate speech Dataset)](https://github.com/jagol/gahd):
+  **CC-BY-4.0**. Citation: Goldzycher, Röttger, Schneider, "Improving
+  Adversarial Data Collection by Supporting Annotators: Lessons from GAHD, a
+  German Hate Speech Dataset", NAACL 2024.
+
+## Datasets used during development (earlier training iterations)
+
+- [GermEval 2018](https://github.com/uds-lsv/GermEval-2018-Data): used under
+  the maintainers' citation terms. Required citation: Wiegand, Siegel,
+  Ruppenhofer, "Overview of the GermEval 2018 Shared Task on the
+  Identification of Offensive Language", Proceedings of GermEval 2018,
+  Vienna.
+- [MLMA hate speech](https://huggingface.co/datasets/nedjmaou/MLMA_hate_speech):
+  **MIT**. Citation: Ousidhoum, Lin, Zhang, Song, Yeung, "Multilingual and
+  Multi-Aspect Hate Speech Analysis", EMNLP 2019.
+
+## Evaluation suites (eval only, never trained on)
+
+- [HateCheck](https://github.com/paul-rottger/hatecheck-data): **CC-BY-4.0**.
+  Citation: Röttger, Vidgen, Nguyen, Waseem, Margetts, Pierrehumbert,
+  "HateCheck: Functional Tests for Hate Speech Detection Models", ACL 2021.
+- [Multilingual HateCheck](https://huggingface.co/Paul/hatecheck): **CC-BY-4.0**.
+  Citation: Röttger, Seelawi, Nozza, Talat, Vidgen, "Multilingual HateCheck:
+  Functional Tests for Multilingual Hate Speech Detection Models", WOAH 2022.
+
+## Generator models (used to create synthetic training data)
+
+- Gemma 4 family ([`google/gemma-4-E4B-it`](https://huggingface.co/google/gemma-4-E4B-it),
+  [`google/gemma-4-12B-it`](https://huggingface.co/google/gemma-4-12B-it)):
+  **Apache-2.0**. Outputs used as synthetic multilingual training text.
+- xAI `grok-4.3` (API): outputs used as synthetic training text for the
+  severity-tail, adversarial and clean-side minimal-pair slices, under xAI's
+  commercial API terms.
+
+No non-commercial or unlicensed datasets are used. None of the above datasets
+are redistributed in this repository.

config.json

@@ -0,0 +1,30 @@
+{
+  "base": "xlm-roberta-base",
+  "content_labels": [
+    "HATEFUL",
+    "HARASSMENT",
+    "THREAT",
+    "SEXUAL",
+    "SELF_HARM"
+  ],
+  "target_labels": [
+    "RACE",
+    "COLOUR",
+    "RELIGION",
+    "DESCENT",
+    "NATIONAL_ETHNIC_ORIGIN",
+    "SEXUAL_ORIENTATION",
+    "GENDER",
+    "DISABILITY",
+    "AGE",
+    "OTHER"
+  ],
+  "max_len": 128,
+  "dropout": 0.1,
+  "hidden_size": 768,
+  "num_layers": 12,
+  "char_dim": 0,
+  "grown_from": "",
+  "vocab_size": 60728,
+  "trimmed_from": "outputs/toxic_a1_s15/final"
+}

labels.json

@@ -0,0 +1,44 @@
+{
+  "content": {
+    "id2label": {
+      "0": "HATEFUL",
+      "1": "HARASSMENT",
+      "2": "THREAT",
+      "3": "SEXUAL",
+      "4": "SELF_HARM"
+    },
+    "label2id": {
+      "HATEFUL": 0,
+      "HARASSMENT": 1,
+      "THREAT": 2,
+      "SEXUAL": 3,
+      "SELF_HARM": 4
+    }
+  },
+  "target": {
+    "id2label": {
+      "0": "RACE",
+      "1": "COLOUR",
+      "2": "RELIGION",
+      "3": "DESCENT",
+      "4": "NATIONAL_ETHNIC_ORIGIN",
+      "5": "SEXUAL_ORIENTATION",
+      "6": "GENDER",
+      "7": "DISABILITY",
+      "8": "AGE",
+      "9": "OTHER"
+    },
+    "label2id": {
+      "RACE": 0,
+      "COLOUR": 1,
+      "RELIGION": 2,
+      "DESCENT": 3,
+      "NATIONAL_ETHNIC_ORIGIN": 4,
+      "SEXUAL_ORIENTATION": 5,
+      "GENDER": 6,
+      "DISABILITY": 7,
+      "AGE": 8,
+      "OTHER": 9
+    }
+  }
+}

tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": true,
+  "backend": "tokenizers",
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "is_local": false,
+  "local_files_only": false,
+  "mask_token": "<mask>",
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "tokenizer_class": "XLMRobertaTokenizer",
+  "unk_token": "<unk>"
+}

toxic.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7700a5ed6772387f89fc17b0c52559b3e8344eec91a0889ebf03623744ea6bad
+size 243

toxic.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cca018426b4e953d028783919caf704e652605a2671ba7470d60d7fece19112a
+size 441

toxic.mlmodelc/metadata.json

@@ -0,0 +1,101 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Mixed (Float16, Palettized (11 bits), Palettized (4 bits), Palettized (9 bits), UInt4)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 5)",
+        "shortDescription" : "",
+        "shape" : "[1, 5]",
+        "name" : "content_logits",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 10)",
+        "shortDescription" : "",
+        "shape" : "[1, 10]",
+        "name" : "target_logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.expandDims" : 3,
+      "Ios18.mul" : 2,
+      "Ios18.realDiv" : 1,
+      "Ios18.scaledDotProductAttention" : 12,
+      "Ios18.greaterEqual" : 2,
+      "Select" : 3,
+      "Tile" : 1,
+      "Ios18.gather" : 2,
+      "Ios18.add" : 29,
+      "Ios18.layerNorm" : 25,
+      "Ios16.reduceSum" : 2,
+      "Ios18.reshape" : 48,
+      "Ios18.constexprLutToDense" : 74,
+      "Ios18.linear" : 74,
+      "Ios18.gelu" : 12,
+      "Ios18.clip" : 1,
+      "Ios18.sub" : 1,
+      "Ios18.cast" : 5,
+      "Ios18.transpose" : 48,
+      "Ios18.notEqual" : 1,
+      "Ios16.cumsum" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-25",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "attention_mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "model_int4",
+    "method" : "predict"
+  }
+]

toxic.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f0b2eb4409d268dcc793f557a57981f38ada83ecfd4f5ec753bf3fd59bf7027
+size 66547284

toxic.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c344bf9a9e4dc4322f0c9deb2e7c014b36b11d45ef1dbacf653a5d4e3ffd7be
+size 136131

toxic.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f0b2eb4409d268dcc793f557a57981f38ada83ecfd4f5ec753bf3fd59bf7027
+size 66547284

toxic.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "A06F1835-64C2-4435-8E00-79E1F5ACFFEF": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "DACCA8AA-FAA6-41E6-9661-A334F8763AFB": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "A06F1835-64C2-4435-8E00-79E1F5ACFFEF"
+}

toxic.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66b0aacd2d65e288b841214620480ee8570143ff0cecdb1a402e86c9bdb826c6
+size 530824007

toxic.tflite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5f64e3c8d3d7abedfcd6ed63ca422e4160ea113a0c38594bd6714647cc1a1d0
+size 75475392

toxic_meta.json

@@ -0,0 +1,146 @@
+{
+  "version": "0.1.0",
+  "languages": [
+    "de",
+    "en",
+    "es",
+    "fr",
+    "it",
+    "nl",
+    "pl",
+    "pt"
+  ],
+  "content_labels": [
+    "HATEFUL",
+    "HARASSMENT",
+    "THREAT",
+    "SEXUAL",
+    "SELF_HARM"
+  ],
+  "target_labels": [
+    "RACE",
+    "COLOUR",
+    "RELIGION",
+    "DESCENT",
+    "NATIONAL_ETHNIC_ORIGIN",
+    "SEXUAL_ORIENTATION",
+    "GENDER",
+    "DISABILITY",
+    "AGE",
+    "OTHER"
+  ],
+  "base_model": "FacebookAI/xlm-roberta-base",
+  "vocab_size": 60728,
+  "recommended": {
+    "threshold": 0.4,
+    "max_length": 128,
+    "per_language_thresholds": {
+      "en": 0.4,
+      "de": 0.55,
+      "fr": 0.45,
+      "nl": 0.55,
+      "pt": 0.6,
+      "es": 0.55,
+      "it": 0.6,
+      "pl": 0.3
+    }
+  },
+  "threshold_menu": {
+    "en": {
+      "tuned": {
+        "threshold": 0.4,
+        "macro_f1": 0.81,
+        "fpr": 0.281
+      },
+      "at_default": {
+        "threshold": 0.4,
+        "macro_f1": 0.81,
+        "fpr": 0.281
+      }
+    },
+    "de": {
+      "tuned": {
+        "threshold": 0.55,
+        "macro_f1": 0.822,
+        "fpr": 0.254
+      },
+      "at_default": {
+        "threshold": 0.4,
+        "macro_f1": 0.803,
+        "fpr": 0.359
+      }
+    },
+    "fr": {
+      "tuned": {
+        "threshold": 0.45,
+        "macro_f1": 0.828,
+        "fpr": 0.247
+      },
+      "at_default": {
+        "threshold": 0.4,
+        "macro_f1": 0.824,
+        "fpr": 0.272
+      }
+    },
+    "nl": {
+      "tuned": {
+        "threshold": 0.55,
+        "macro_f1": 0.768,
+        "fpr": 0.238
+      },
+      "at_default": {
+        "threshold": 0.4,
+        "macro_f1": 0.764,
+        "fpr": 0.298
+      }
+    },
+    "pt": {
+      "tuned": {
+        "threshold": 0.6,
+        "macro_f1": 0.8,
+        "fpr": 0.203
+      },
+      "at_default": {
+        "threshold": 0.4,
+        "macro_f1": 0.791,
+        "fpr": 0.288
+      }
+    },
+    "es": {
+      "tuned": {
+        "threshold": 0.55,
+        "macro_f1": 0.803,
+        "fpr": 0.207
+      },
+      "at_default": {
+        "threshold": 0.4,
+        "macro_f1": 0.796,
+        "fpr": 0.272
+      }
+    },
+    "it": {
+      "tuned": {
+        "threshold": 0.6,
+        "macro_f1": 0.779,
+        "fpr": 0.234
+      },
+      "at_default": {
+        "threshold": 0.4,
+        "macro_f1": 0.775,
+        "fpr": 0.309
+      }
+    },
+    "pl": {
+      "tuned": {
+        "threshold": 0.3,
+        "macro_f1": 0.759,
+        "fpr": 0.305
+      },
+      "at_default": {
+        "threshold": 0.4,
+        "macro_f1": 0.751,
+        "fpr": 0.271
+      }
+    }
+  }
+}

toxic_tokenizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a1e9c02fb26535f9293704a3d9b8f0a1c5a8178ed0db807d9d7c07caed0ead0
+size 772537