Update app.py

app.py

@@ -1,23 +1,32 @@
 import gradio as gr
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
 import json
 import html
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
 
-# ---------- 使用 Phi-3-mini (3.8B，极快且稳定) ----------
-MODEL_NAME = "microsoft/Phi-3-mini-4k-instruct"
+# ---------- 模型：Qwen2.5-3B-Instruct (4bit 量化，稳定且快) ----------
+MODEL_NAME = "Qwen/Qwen2.5-3B-Instruct"
 
-print("🔄 加载轻量模型 Phi-3-mini...")
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.float16,
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_quant_type="nf4"
+)
+
+print("🔄 加载 Qwen2.5-3B 模型 (4bit)...")
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
 model = AutoModelForCausalLM.from_pretrained(
     MODEL_NAME,
-    torch_dtype=torch.float16,
+    quantization_config=quantization_config,
     device_map="auto",
-    trust_remote_code=True
+    trust_remote_code=True,
+    torch_dtype=torch.float16
 )
 print("✅ 模型加载完成")
 
-SYSTEM_PROMPT = """你是一个英语精读辅助工具。从用户提供的英文文本中提取 8~20 个值得学习的表达（短语、搭配、句式，不是单词）。
+# ---------- 提示词 ----------
+SYSTEM_PROMPT = """你是一个英语精读辅助工具。从用户提供的英文文本中提取 8~20 个值得学习的表达（短语、搭配、句式，不是单个单词）。
 
 对每个表达，输出一个 JSON 对象，包含：
 - "expression": 原文中的表达
@@ -42,50 +51,52 @@ def parse_output(raw):
         data = json.loads(json_str)
         required = {"expression","meaning","explanation","original_context","extra_example"}
         return [item for item in data if isinstance(item,dict) and required.issubset(item.keys())]
-    except:
+    except Exception as e:
+        print("解析错误:", e)
         return []
 
 def generate_cards(exprs):
     if not exprs:
-        return '<div style="padding:2rem;text-align:center">⚠️ 未提取到表达，请尝试不同文本</div>'
+        return '<div style="padding:2rem;text-align:center;color:#666;">⚠️ 未提取到表达，请尝试不同文本。</div>'
     cards = ""
     for e in exprs:
         cards += f"""
-        <div style="background:white;border-radius:20px;border:1px solid #e0e0e0;padding:1.2rem;margin-bottom:1rem;">
-            <div style="font-weight:600;font-size:1.2rem;margin-bottom:0.8rem;">{html.escape(e['expression'])}</div>
-            <div><strong>Meaning</strong><br>{html.escape(e['meaning'])}</div>
-            <div style="margin-top:0.5rem;"><strong>Explanation</strong><br>{html.escape(e['explanation'])}</div>
-            <div style="margin-top:0.5rem;"><strong>Original Context</strong><br>{html.escape(e['original_context'])}</div>
-            <div style="margin-top:0.5rem;"><strong>Extra Example</strong><br>{html.escape(e['extra_example'])}</div>
+        <div style="background:white;border-radius:20px;border:1px solid #e2e2e0;padding:1.2rem;margin-bottom:1rem;box-shadow:0 1px 2px rgba(0,0,0,0.02);">
+            <div style="font-weight:600;font-size:1.2rem;margin-bottom:0.8rem;color:#1a1a1a;">{html.escape(e['expression'])}</div>
+            <div><strong style="color:#555;">Meaning</strong><br><span style="color:#333;">{html.escape(e['meaning'])}</span></div>
+            <div style="margin-top:0.6rem;"><strong style="color:#555;">Explanation</strong><br><span style="color:#333;">{html.escape(e['explanation'])}</span></div>
+            <div style="margin-top:0.6rem;"><strong style="color:#555;">Original Context</strong><br><span style="color:#333;">{html.escape(e['original_context'])}</span></div>
+            <div style="margin-top:0.6rem;"><strong style="color:#555;">Additional Example</strong><br><span style="color:#333;">{html.escape(e['extra_example'])}</span></div>
         </div>
         """
-    return f'<div style="max-width:900px;margin:0 auto">{cards}</div>'
+    return f'<div style="max-width:950px;margin:0 auto">{cards}</div>'
 
 def analyze(text, progress=gr.Progress()):
     if len(text.strip()) < 20:
-        return "<div style='color:red'>请至少输入20个字符</div>", 0
+        return "<div style='color:#d14;padding:1rem;'>⚠️ 请至少输入20个字符的英文文本。</div>", 0
     messages = [
         {"role": "system", "content": SYSTEM_PROMPT},
         {"role": "user", "content": USER_PROMPT_TEMPLATE.format(text=text)}
     ]
-    inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
-    progress(0.4, desc="模型推理中...")
+    text_input = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = tokenizer(text_input, return_tensors="pt").to(model.device)
+    progress(0.4, desc="模型推理中（约20-30秒）...")
     with torch.no_grad():
         outputs = model.generate(
-            inputs,
-            max_new_tokens=1024,
-            do_sample=False,       # 贪心解码，最快
+            **inputs,
+            max_new_tokens=1536,
+            do_sample=False,
             temperature=1.0,
             pad_token_id=tokenizer.eos_token_id
         )
-    response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
-    exprs = parse_output(response)
+    generated = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
+    exprs = parse_output(generated)
     if not exprs:
-        return "<div style='color:red'>解析失败，请重试</div>", 0
+        return "<div style='color:#d14;padding:1rem;'>⚠️ 解析失败，请重试或更换文本。</div>", 0
     return generate_cards(exprs), len(exprs)
 
-# ---------- 极简浅色主题（不用任何复杂 CSS，确保生效）----------
-LIGHT_THEME = gr.themes.Soft(
+# ---------- 强制浅色主题 ----------
+theme = gr.themes.Soft(
     primary_hue="neutral",
     secondary_hue="neutral",
     font=gr.themes.GoogleFont("Inter"),
@@ -96,22 +107,26 @@ LIGHT_THEME = gr.themes.Soft(
     block_background_fill="white",
     block_border_width="1px",
     block_border_color="#e2e2e0",
+    block_title_background_fill="white",
+    block_label_background_fill="white",
+    input_background_fill="white",
+    input_border_color="#e2e2e0",
 )
 
-with gr.Blocks(theme=LIGHT_THEME, title="InContext") as demo:
-    gr.Markdown("# InContext\nLearn English Expressions Through Real Content")
+with gr.Blocks(theme=theme, title="InContext — Learn English from Real Content", css="footer {visibility: hidden}") as demo:
+    gr.Markdown("# InContext\n### Learn English Expressions Through Real Content")
     with gr.Row():
-        textbox = gr.Textbox(lines=10, placeholder="Paste English content here...", label="")
+        textbox = gr.Textbox(lines=10, placeholder="Paste any English content here...\n\nExample:\nThese links and codes are for registered Build Small participants only.\nPlease don't forward or post them publicly.\nRedemptions are capped.\nEvery leaked claim takes compute away from a fellow builder.", label="")
     btn = gr.Button("Analyze", variant="primary")
-    header = gr.HTML('<div style="display:flex;justify-content:space-between;margin:1rem 0"><h3>Expressions Found</h3><span id="count">—</span></div>')
+    header = gr.HTML('<div style="display:flex;justify-content:space-between;align-items:baseline;margin:1.5rem 0 1rem 0;"><h3 style="margin:0;">Expressions Found</h3><span id="count" style="background:#efefec;padding:0.2rem 0.7rem;border-radius:30px;font-size:0.8rem;">—</span></div>')
     output = gr.HTML()
     count_state = gr.Number(visible=False)
-    
+
     def run(text):
         html, cnt = analyze(text)
-        new_header = f'<div style="display:flex;justify-content:space-between;margin:1rem 0"><h3>Expressions Found</h3><span>{cnt}</span></div>'
+        new_header = f'<div style="display:flex;justify-content:space-between;align-items:baseline;margin:1.5rem 0 1rem 0;"><h3 style="margin:0;">Expressions Found</h3><span style="background:#efefec;padding:0.2rem 0.7rem;border-radius:30px;font-size:0.8rem;">{cnt}</span></div>'
         return new_header, html, cnt
-    
+
     btn.click(run, [textbox], [header, output, count_state])
 
 demo.queue().launch()