Upload handler.py with huggingface_hub

handler.py

@@ -52,13 +52,14 @@ class EndpointHandler:
 
         if is_custom_prompt:
             # Custom prompt format for unsupported languages
-            # Use Gemma chat format: <start_of_turn>user\n{instruction} {text}<end_of_turn>\n<start_of_turn>model\n
-            prompt = f"<start_of_turn>user\n{target_lang} {text}<end_of_turn>\n<start_of_turn>model\n"
-            inputs = self.processor.tokenizer(
+            # Add explicit instruction to return ONLY the translation
+            prompt = f"<start_of_turn>user\n{target_lang} Output only the translation, no explanations.\n\n{text}<end_of_turn>\n<start_of_turn>model\n"
+            tokenized = self.processor.tokenizer(
                 prompt,
                 return_tensors="pt",
                 add_special_tokens=True
-            ).to(self.model.device)
+            )
+            inputs = {k: v.to(self.model.device) for k, v in tokenized.items()}
         else:
             # Standard language code: use structured message format
             messages = [