Skeleton ShieldGemma class

keras_hub/src/models/gemma/gemma_presets.py

@@ -206,7 +206,7 @@
         "metadata": {
             "description": "2 billion parameter, 26-layer, ShieldGemma model.",
             "params": 2614341888,
-            "official_name": "Gemma",
+            "official_name": "ShieldGemma",
             "path": "gemma",
             "model_card": "https://www.kaggle.com/models/google/shieldgemma",
         },
@@ -216,7 +216,7 @@
         "metadata": {
             "description": "9 billion parameter, 42-layer, ShieldGemma model.",
             "params": 9241705984,
-            "official_name": "Gemma",
+            "official_name": "ShieldGemma",
             "path": "gemma",
             "model_card": "https://www.kaggle.com/models/google/shieldgemma",
         },
@@ -226,7 +226,7 @@
         "metadata": {
             "description": "27 billion parameter, 42-layer, ShieldGemma model.",
             "params": 27227128320,
-            "official_name": "Gemma",
+            "official_name": "ShieldGemma",
             "path": "gemma",
             "model_card": "https://www.kaggle.com/models/google/shieldgemma",
         },

keras_hub/src/models/gemma/shieldgemma.py

@@ -0,0 +1,80 @@
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.gemma import gemma_causal_lm
+from keras_hub.src.models.task import Task
+
+
+class ShieldGemmaViolationProbaility(keras.layers.Layer):
+    """Relative probabilities for the 'Yes' (violating) and 'No' tokens."""
+
+    def __init__(self, yes_token_idx, no_token_idx, **kw):
+        super().__init__(**kw)
+        self.yes_token_idx = yes_token_idx
+        self.no_token_idx = no_token_idx
+
+    def call(self, logits, padding_mask):
+        last_prompt_index = keras.ops.cast(
+            keras.ops.sum(padding_mask, axis=1) - 1, "int32"
+        )
+        last_logits = keras.ops.take(logits, last_prompt_index, axis=1)[:, 0]
+        yes_logits = last_logits[:, self.yes_token_idx]
+        no_logits = last_logits[:, self.no_token_idx]
+        yes_no_logits = keras.ops.stack((yes_logits, no_logits), axis=1)
+        return keras.ops.softmax(yes_no_logits, axis=1)
+
+
+@keras_hub_export("keras_hub.models.ShieldGemma")
+class ShieldGemma(Task):
+    """A ShieldGemma model for safety content moderation, built on Gemma 2.
+
+    ShieldGemma is a Gemma 2 variant fine-tuned to detect and predict violations
+    of four harm types—Harrassment, Hate Speech, Dangerous Content, and
+    Sexual Content—in text content from a user or model. Architecturally,
+    the weights are the same as any other Gemma 2 class, but the prediction is
+    augmented with a final layer that returns the probability that the provided
+    content violates the harm type specified in the prompt. The probability is
+    computed as the relative probabilities of the `Yes` (violating) and `No`
+    (non-violating) tokens at the final prompt token, i.e., is the next most
+    likley token a yes or a no.
+
+    Links:
+
+    *   https://arxiv.org/abs/2407.21772
+    *   https://ai.google.dev/gemma/docs/shieldgemma/model_card
+    *   https://ai.google.dev/responsible/docs/safeguards/shieldgemma
+    *   https://www.kaggle.com/models/google/shieldgemma
+
+    Args:
+        gemma: A `keras_hub.models.GemmaCausalLM` initialized with ShieldGemma
+            weights.
+
+    Examples:
+
+    Coming soon.
+    """
+
+    backbone_cls = gemma_causal_lm.GemmaCausalLM.backbone_cls
+    preprocessor_cls = gemma_causal_lm.GemmaCausalLM.preprocessor_cls
+
+    def __init__(self, gemma: gemma_causal_lm.GemmaCausalLM, **kwargs):
+        # === Layers ===
+        self.gemma = gemma
+        self.backbone = self.gemma.backbone
+        self.preprocessor = self.gemma.preprocessor
+        self.yes_no_layer = ShieldGemmaViolationProbaility(
+            yes_token_idx=self.preprocessor.tokenizer.token_to_id("Yes"),
+            no_token_idx=self.preprocessor.tokenizer.token_to_id("No"),
+        )
+
+        # === Functional Model ===
+        inputs = self.gemma.input
+        logits = self.gemma(inputs)
+        outputs = self.yes_no_layer(logits, inputs["padding_mask"])
+        super().__init__(inputs=inputs, outputs=outputs, **kwargs)
+
+    @classmethod
+    def from_preset(cls, **kwargs):
+        """Instantiate a `keras_hub.models.ShieldGemma` from a model preset."""
+        gemma = gemma_causal_lm.GemmaCausalLM.from_preset(**kwargs)
+        return cls(gemma)