Update to routangseng-voice-4b with matching system prompt and UI

README.md

@@ -1,16 +1,19 @@
 ---
-title: routangseng-chat-gpu
+title: 肉糖生 Chat GPU
 emoji: 🥩
-colorFrom: indigo
-colorTo: pink
+colorFrom: red
+colorTo: purple
 sdk: gradio
-sdk_version: 5.46.1
-python_version: 3.10.13
+sdk_version: "5.23.0"
 app_file: app.py
 pinned: false
+license: apache-2.0
 suggested_hardware: zero-a10g
 ---
 
 # 肉糖生 Chat GPU
 
-Gradio + ZeroGPU + Transformers runtime bootstrap for `bobber/routangseng-qwen35-0.8b-abliterated`.
+Qwen3.5-4B voice LoRA fine-tuned model running on ZeroGPU.
+
+- Model: `bobber/routangseng-voice-4b`
+- Style: 结论先行，不和稀泥

app.py

@@ -9,9 +9,9 @@ import gradio as gr
 import spaces
 import torch
 
-MODEL_ID = "bobber/routangseng-qwen35-0.8b-abliterated"
-SYSTEM_PROMPT = "你是一个中文内容创作者，表达理性、结构化、接地气，先讲结论再展开分析，并保持多轮对话一致性。"
-MAX_TOKENS = 512
+MODEL_ID = "bobber/routangseng-voice-4b"
+SYSTEM_PROMPT = "你是肉糖生，一个接地气的中文时政分析者。风格：结论先行，再用结构化分析展开；敢于质疑主流叙事，不和稀泥；用类比和现实例子把复杂问题讲透；语气直率但逻辑严密。回答时先给核心判断，再分层拆解，最后给出预测或建议。"
+MAX_TOKENS = 1024
 TEMPERATURE = 0.7
 TOP_P = 0.9
 
@@ -103,7 +103,7 @@ def load_model():
         _tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=False)
         _model = AutoModelForCausalLM.from_pretrained(
             MODEL_ID,
-            torch_dtype=torch.float16,
+            torch_dtype=torch.bfloat16,
             device_map="auto",
             trust_remote_code=False,
         )
@@ -148,26 +148,29 @@ def chat_fn(message, history):
     thread.join()
 
 
-with gr.Blocks(theme=gr.themes.Soft()) as demo:
+CSS = """
+.gradio-container { max-width: 900px !important; }
+footer { display: none !important; }
+"""
+
+with gr.Blocks(theme=gr.themes.Soft(), css=CSS) as demo:
     gr.Markdown(
         """
 # 🥩 肉糖生 Chat
 
-Qwen3.5-0.8B SFT · ZeroGPU · Transformers
+**Qwen3.5-4B Voice LoRA** · ZeroGPU · 结论先行，不和稀泥
+
+<small>基于 `bobber/routangseng-voice-4b` · 首次启动需安装最新 Transformers（约30秒）</small>
 """
     )
-    gr.Markdown(
-        "首次启动可能会慢一些：Space 需要在运行时引导安装最新 Transformers 主线版本，以获得 Qwen3.5 支持。"
-    )
     gr.ChatInterface(
         fn=chat_fn,
         type="messages",
-        title="肉糖生 · Qwen3.5-0.8B Abliterated SFT",
-        description="基于 bobber/routangseng-qwen35-0.8b-abliterated",
         examples=[
-            "用三句话介绍你自己。",
-            "帮我写一段中文产品介绍，风格理性、接地气。",
-            "怎么理解机器人+建筑行业的结合？",
+            "白领工作都被AI不断代替，现在学生还在用传统方式积累白领知识，这不是学了个寂寞嘛？",
+            "为什么很多国家的年轻人不想生孩子？这个趋势能逆转吗？",
+            "中美关系未来五年会怎么走？从结构性矛盾的角度讲讲。",
+            "怎么理解"小共同体"和"大共同体"的关系？",
         ],
         fill_height=True,
     )

requirements.txt

@@ -1,11 +1,4 @@
-accelerate>=1.0.0
-regex
-tokenizers
-safetensors
-sentencepiece
-filelock
-packaging
-requests
-pyyaml
-numpy
-tqdm
+gradio>=4.0
+torch
+spaces
+peft