Update app.py

app.py

@@ -1,48 +1,31 @@
-import streamlit as st
-from transformers import pipeline
-
-# Page setup
-st.set_page_config(page_title="Hate Speech Detector", page_icon="🚨", layout="centered")
-st.title("🚨 Hate Speech Classification Demo")
-st.markdown("""
-This app uses the fine-tuned model `JenniferHJF/qwen1.5-emoji-finetuned` to predict whether the given input text contains **offensive or hateful content**.
-
-The model returns:
-- `1` if the content is offensive
-- `0` if it is not offensive
-
-⚠️ Note: This is a demo and the model may not be perfect in detecting nuanced or implicit hate speech.
-""")
-
-# Example inputs
-examples = [
-    "You're a disgrace to this country.",
-    "Hope you have a great day!",
-    "Why are you even alive?",
-    "That was really rude and uncalled for.",
-    "You are amazing and smart!",
-    "Get lost, nobody wants you here."
-]
-
-selected_example = st.selectbox("📘 Choose an example sentence:", options=examples, index=0)
-text = st.text_area("📝 Or enter your own text below:", value=selected_example, height=150)
-
-if st.button("🚀 Analyze"):
-    with st.spinner("Running model inference..."):
-        classifier = pipeline("text-generation", model="JenniferHJF/qwen1.5-emoji-finetuned", max_new_tokens=20)
-        output = classifier(f"""Please determine whether the following text is offensive.
-Reply with '1' for offensive, '0' for non-offensive.
-Text: {text}
-""")[0]["generated_text"]
-
-        # Extract the last '0' or '1' from output
-        prediction = "Unknown"
-        if "1" in output.strip().splitlines()[-1]:
-            prediction = "Offensive (1)"
-        elif "0" in output.strip().splitlines()[-1]:
-            prediction = "Non-Offensive (0)"
-
-        st.markdown(f"### ✅ Prediction: `{prediction}`")
-        st.code(output.strip(), language="text")
-else:
-    st.info("👈 Enter text and click 'Analyze' to begin.")
+from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
+
+# Load Qwen 微调模型用于 emoji 转换
+emoji_translator = pipeline(
+    "text-generation", 
+    model="JenniferHJF/qwen1.5-emoji-finetuned", 
+    tokenizer="JenniferHJF/qwen1.5-emoji-finetuned",
+    max_new_tokens=20,
+    trust_remote_code=True
+)
+
+# Load zero-shot/offensive-classification model（可替换为 ChatGLM3、DeepSeek 等）
+offensive_classifier = pipeline(
+    "text-classification",
+    model="s-nlp/roberta-offensive-language-detection"  # 示例模型，可换大模型
+)
+
+# Unified prediction function
+def classify_text_with_emoji(raw_text):
+    # Step 1: Convert emojis ➝ Chinese
+    prompt = f"输入：{raw_text}\n输出："
+    converted = emoji_translator(prompt)[0]['generated_text']
+    # 拿最后一行当输出结果（避免生成前缀）
+    translated_text = converted.strip().splitlines()[-1]
+    
+    # Step 2: Run classification
+    result = offensive_classifier(translated_text)[0]
+    label = result['label']
+    score = result['score']
+
+    return translated_text, label, score