Publish toxicity model and artifacts

README.md

@@ -31,23 +31,29 @@ by [HikmaAI](https://huggingface.co/HikmaAI).
 ## Model Description
 
 - **Task**: Binary classification (benign=0, toxic=1)
+- **Version**: 3.0.0
 - **Base model**: `xlm-roberta-large`
 - **Languages**: 11 (en, vi, hi, th, zh, ja, ru, ar, sv, es, it)
 - **Export formats**: ONNX FP32 + FP16 + INT8 (x86-safe dynamic)
 
+Beyond declarative toxicity (slurs, hate speech, threats *in the words*), the
+classifier is trained to flag politely-phrased harmful **requests**: harm
+solicitation, doxxing/privacy requests, solicited discrimination, atrocity
+denial, and illicit-capability questions.
+
 ## Performance
 
 
 | Metric | Score |
 |--------|-------|
-| loss | 0.3252 |
-| accuracy | 0.9415 |
-| precision | 0.9133 |
-| recall | 0.9234 |
-| f1 | 0.9183 |
+| loss | 0.1001 |
+| accuracy | 0.9834 |
+| precision | 0.9507 |
+| recall | 0.8914 |
+| f1 | 0.9201 |
 
 
-Optimized threshold: **0.5000**
+Optimized threshold: **0.6600** (val recall: 0.8854)
 
 ## Usage (ONNX)
 
@@ -73,9 +79,10 @@ outputs = model(**inputs)
 
 - Epochs: 5
 - Learning rate: 2e-05
-- Batch size: 8
+- Batch size: 32
 - Class weights: [1.0, 1.0]
-- Dataset: multilingual (11 languages), 12+ sources + synthetic data
+- Dataset: multilingual (11 languages), ~90/10 benign/toxic, 12+ sources +
+  targeted synthetic data (request-style harm cells in all 11 languages)
 
 ## License
 

model_card.json

@@ -1,17 +1,17 @@
 {
   "classifier_name": "toxicity",
-  "version": "2.0.0",
+  "version": "3.0.0",
   "task_type": "sequence_classification",
   "base_model": "xlm-roberta-large",
   "metrics": {
-    "loss": 0.325167179107666,
-    "accuracy": 0.9415098374679213,
-    "precision": 0.9133016627078385,
-    "recall": 0.9234464124887422,
-    "f1": 0.9183460217942977
+    "loss": 0.10005442798137665,
+    "accuracy": 0.9834395959261406,
+    "precision": 0.9507429829389102,
+    "recall": 0.8913828689370485,
+    "f1": 0.9201065246338216
   },
   "dataset_hash": "",
-  "created_at": "2026-07-23T11:06:52.180593Z",
+  "created_at": "2026-07-28T08:45:42.899160Z",
   "export_format": "onnx",
   "quantization": "fp16+int8"
 }

onnx/fp16/model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4b15b3eac7c859f1559d23a12a82ddef60759d9d4a96dbf450929a50125b9fc0
+oid sha256:c4ae1390fc655d08396be5bce51243c7677cccdde86c8bc1264f11d2090733a9
 size 492775

onnx/fp16/model.onnx_data

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:81b80b32a9357ce623f13b6e2917d2c2a89e68762e889a85e951df01e14cdd72
+oid sha256:b798d973d2d6dbab396980ebdfa268eb6d37726e7d378b4d18e7c719318ed407
 size 1119784960

onnx/fp32/model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:18e97156bb4bf4343f454ba6116999c6c3d840ae26894090e04cd906f9bfa960
+oid sha256:7ff5a98a0d0fecd92272433f9bd8b7746aae77b5bb60973426cbf1f282439883
 size 489246

onnx/fp32/model.onnx_data

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:be934f74c8dbfeb3a8fd62c19590cc0203bdaacd1df925a1628b7106ae1a3ac3
+oid sha256:ff30e88059c46c124df879265eec373a2cc63ba8807a8c50444c5072e18be7c2
 size 2239574032

onnx/int8/model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aa9ea6c4b572ee919d975b03f8fd09bf9ee887adece704217920cc2a56f40a12
-size 562892984
+oid sha256:b66e23667b483e8dc9f200586583753297ad65332d1fc3460bae91da6816dae6
+size 562892985

threshold.json

@@ -1,6 +1,7 @@
 {
-  "threshold": 0.5,
-  "objective": "production-robust (bimodal model; F1 flat 0.045-0.9, 0.5 = max margin, gateway-aligned)",
-  "val_f1": 0.9128,
-  "note": "max-F1 sweep lands at 0.045 but that is the low edge of a narrow plateau; 0.5 chosen for drift robustness"
+  "threshold": 0.66,
+  "val_recall": 0.8853743470690656,
+  "val_precision": 0.9442896935933147,
+  "val_f1": 0.9138834806050622,
+  "val_accuracy": 0.9837501766285149
 }