Initial release of InferenceGuard RoBERTa risk classifier

README.md

@@ -0,0 +1,41 @@
+---
+language:
+- en
+license: mit
+tags:
+- privacy
+- inferential-privacy
+- roberta
+- multi-task-classification
+- llm-firewall
+base_model: roberta-base
+pipeline_tag: text-classification
+---
+
+# InferenceGuard RoBERTa Multi-Task Risk Classifier
+
+Production multi-task classifier designed to detect and quantify inferential privacy risks in LLM prompts.
+The model predicts inferability across four demographic dimensions: age, location, occupation, and education.
+It acts as the primary risk estimation engine within the InferenceGuard privacy firewall framework.
+
+## Model Details
+
+- Backbone: roberta-base
+- Task: Multi-task binary inferability classification
+- Target Attributes: age, location, occupation, education
+- Calibration: Attribute-level temperature scaling
+
+## Quick Start
+
+```python
+from transformers import AutoTokenizer
+from src.risk_model.roberta_model import RobertaRiskClassifier
+
+repo_id = "AT707/InferenceGuard-Roberta-Risk"
+tokenizer = AutoTokenizer.from_pretrained(repo_id)
+model = RobertaRiskClassifier.from_pretrained(repo_id)
+
+inputs = tokenizer("I graduated from MIT in 2012 and work as a surgeon in Boston.", return_tensors="pt")
+probs = model.predict_proba(**inputs)
+print(probs)
+```

config.json

@@ -0,0 +1,38 @@
+{
+  "transformers_version": "5.22.1",
+  "architectures": null,
+  "output_hidden_states": false,
+  "return_dict": true,
+  "dtype": null,
+  "chunk_size_feed_forward": 0,
+  "is_encoder_decoder": false,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "problem_type": null,
+  "_name_or_path": "",
+  "base_model_name": "roberta-base",
+  "num_heads": 4,
+  "num_classes_per_head": 1,
+  "hidden_size": 768,
+  "classifier_dropout": 0.1,
+  "temperature_values": [
+    1.0,
+    1.0,
+    1.0,
+    1.0
+  ],
+  "attributes": [
+    "age",
+    "location",
+    "occupation",
+    "education"
+  ],
+  "model_type": "inferenceguard-roberta",
+  "output_attentions": false
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33eef435c3304196143c693ef61aba400cc0fbd14a9fa7ad92b724726e51c223
+size 503401423

tokenizer_config.json

@@ -0,0 +1,17 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "mask_token": "<mask>",
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": "<unk>"
+}