Spaces:

dqy08
/

InfoRadar

Running

App Files Files Community

dqy08 commited on Jan 29

Commit

c6aeab1

1 Parent(s): cbb60f2

更新Dockerfile以使用Qwen3-0.6B模型；更新README和HTML文件；更新示例数据以匹配新模型。

Browse files

Files changed (19) hide show

Dockerfile +3 -1
README.md +9 -11
README.zh-CN.md +0 -10
backend/language_checker.py +18 -12
backend/runtime_config.py +1 -1
client/src/content/home.en.html +3 -2
client/src/content/home.zh.html +1 -1
client/src/index.html +3 -1
data/demo/public/CN/GPT-2 large unicorn text(中文翻译).json +0 -0
data/demo/public/CN/GPT-2 small top_k 40 temp .7 (中文翻译).json +0 -0
data/demo/public/CN/GPT-2 small top_k 5 temp 1 (中文翻译).json +0 -0
data/demo/public/CN/human_ NYTimes article (中文翻译).json +0 -0
data/demo/public/CN/human_ academic text (中文翻译).json +0 -0
data/demo/public/GPT-2 large unicorn text +0 -0
data/demo/public/GPT-2 small top_k 5 temp 1.json +911 -0
data/demo/public/Wiki - Cristiano Ronaldo.json +0 -0
data/demo/public/human_ NYTimes article.json +0 -0
data/demo/public/human_ academic text.json +0 -0
data/demo/public/human_ woodchuck.json +964 -0

Dockerfile CHANGED Viewed

@@ -55,4 +55,6 @@ ENV FORCE_INT8=1
 EXPOSE 7860
-CMD ["python", "server.py", "--model", "qwen3.0-14b", "--address", "0.0.0.0", "--port", "7860"]

 EXPOSE 7860
+# CMD ["python", "server.py", "--model", "qwen3.0-14b", "--address", "0.0.0.0", "--port", "7860"]
+CMD ["python", "server.py", "--model", "qwen3.0-0.6b", "--address", "0.0.0.0", "--port", "7860"]
+ENV FORCE_INT8=0

README.md CHANGED Viewed

@@ -1,9 +1,10 @@
 ---
-title: InfoRadar
 emoji: 📡
 colorFrom: blue
-colorTo: indigo
 sdk: docker
 app_port: 7860
 pinned: false
 license: apache-2.0
@@ -13,24 +14,21 @@ license: apache-2.0
 # InfoRadar (Information Radar)
-**InfoRadar** is a visual tool for **analyzing text information density and efficiency**.
-Unlike traditional "AI detectors" that simply classify text as "Human vs. AI", InfoRadar focuses on evaluating the **quality of information**. By visualizing "surprisal" (information content), it intuitively reveals information flow patterns, helping to identify low-information "nonsense" (whether AI hallucinations or human verbosity) and highlighting high-density core information.
 ## 🚀 Core Features
--   **Information Density Visualization**: Color-coded analysis based on token-level surprisal (`-log p`).
-    -   ⚪ **Transparent**: High predictability (Low information / Common phrases / "Filler")
-    -   🔴 **Red**: High information content (Surprising / Specific / "Core Content")
 ## 💡 Tribute
 InfoRadar is engineered based on the classic project [GLTR.io](http://gltr.io) developed by Hendrik Strobelt et al. in 2019. GLTR was a web demo that pioneered the use of GPT-2 prediction probabilities to detect generated text.
-The difference lies in the goal of this project: **not to "detect AI text", but to "evaluate text quality"**:
-1.  **From "Detection" to "Evaluation"**: Shifting focus from "Is this written by AI?" to "Is this content efficient and valuable?"
-2.  **Information Theoretic Perspective**: Introducing cognitive linguistics concepts (such as Surprisal Theory, UID) to measure text quality from first principles.
 ## 📦 Quick Start

 ---
+title: InfoRadar – Visualize Text Information Density
 emoji: 📡
 colorFrom: blue
+colorTo: red
 sdk: docker
+short_description: analyzes text to visualize token-level information density
 app_port: 7860
 pinned: false
 license: apache-2.0
 # InfoRadar (Information Radar)
+Tired of low-quality articles? Struggling to find key points in long texts? Want to skip redundancy and fluff at a glance? Or just curious about the information-theoretic nature of language?
+**Try InfoRadar.** It uses large language models to analyze text information density and visualizes where the important parts are. The color intensity of each character indicates how much information it carries.
 ## 🚀 Core Features
+-   **Information Density Visualization**: Color-coded analysis based on token-level surprisal (`-log₂ p`).
+    -   ⚪ **Transparent**: High predictability (low information / common phrases / filler)
+    -   🔴 **Red**: High information content (surprising / specific / core content)
 ## 💡 Tribute
 InfoRadar is engineered based on the classic project [GLTR.io](http://gltr.io) developed by Hendrik Strobelt et al. in 2019. GLTR was a web demo that pioneered the use of GPT-2 prediction probabilities to detect generated text.
+The difference lies in the goal: **not to "detect AI text", but to "evaluate text quality"**. When we dislike AI text, we actually dislike low-quality text; the key is information quality. InfoRadar focuses on "information quality" rather than "AI signs", though it can help spot AI-generated nonsense with no information content. Currently **Qwen3-14B-Base** is used for analysis.
 ## 📦 Quick Start

README.zh-CN.md CHANGED Viewed

@@ -1,13 +1,3 @@
----
-title: InfoRadar
-emoji: 📡
-colorFrom: blue
-colorTo: indigo
-sdk: docker
-app_port: 7860
-license: apache-2.0
----
 **[English](README.md)** | 简体中文
 # InfoRadar (信息雷达)












1	[English](README.md) \| 简体中文
2
3	# InfoRadar (信息雷达)

backend/language_checker.py CHANGED Viewed

@@ -120,11 +120,11 @@ class AbstractLanguageChecker:
         获取计算设备
         优先级：
-          1. 显式强制 CPU（FORCE_CPU 环境变量）
           2. 自动检测最优设备（cuda > mps > cpu）
         """
         # 如果显式要求 CPU，直接返回（唯一有意义的强制场景）
-        if os.environ.get('FORCE_CPU'):
             return torch.device("cpu")
         # 自动选择最优设备
@@ -210,10 +210,10 @@ class QwenLM(AbstractLanguageChecker):
         load_description = "模型"
         # 环境变量配置
-        # FORCE_INT8: 启用 INT8 量化（适用于 CPU 和 CUDA，实验性，在某些情况下会降低性能）
-        # CPU_FORCE_BFLOAT16: 启用 bfloat16（仅适用于 CPU，需硬件加速支持，否则会降低性能）
-        force_int8 = os.environ.get('FORCE_INT8')
-        force_bfloat16 = os.environ.get('CPU_FORCE_BFLOAT16')
         # 检测是否为 AWQ 模型（自动检测）
         is_awq_model = self._is_awq_model(model_path)
@@ -239,12 +239,12 @@ class QwenLM(AbstractLanguageChecker):
                 use_int8 = True
                 device_map = "cpu"
                 load_description = "模型（INT8量化）"
-                print("⚠️  启用 INT8 量化（实验性，在某些情况下会降低性能）")
             elif force_bfloat16:
                 dtype = torch.bfloat16
                 use_low_cpu_mem = True
-                print("⚠️  启用 bfloat16（需硬件加速支持，否则会降低性能）")
             else:
                 # 默认: float32
@@ -260,7 +260,7 @@ class QwenLM(AbstractLanguageChecker):
             if force_int8:
                 use_int8 = True
                 load_description = "模型（INT8量化）"
-                print("⚠️  启用 INT8 量化")
             else:
                 dtype = torch.float16
                 print("🔧 dtype: float16")
@@ -271,7 +271,7 @@ class QwenLM(AbstractLanguageChecker):
             print(f"🔧 {self.device.type.upper()} 模式：自动设备分配")
             if force_int8:
-                print("⚠️  MPS 不支持 INT8 量化，已忽略 FORCE_INT8 环境变量")
             device_map = "auto"
             dtype = torch.float16
@@ -351,6 +351,9 @@ class QwenLM(AbstractLanguageChecker):
         device_name = DeviceManager.get_device_name(self.device)
         print(f"✓ {model_display_name} 模型已加载 ({device_name})")
     def _load_model_with_int8_cuda(
         self,
@@ -787,8 +790,11 @@ class QwenLM(AbstractLanguageChecker):
             DeviceManager.clear_cache(self.device)
             gc.collect()
-            # 打印分析任务完成后的内存统计
-            if self.device.type == "cuda":
                 device_idx = self.device.index if self.device.index is not None else 0
                 DeviceManager.print_cuda_memory_summary(device=device_idx)

         获取计算设备
         优先级：
+          1. 显式强制 CPU（FORCE_CPU=1 环境变量）
           2. 自动检测最优设备（cuda > mps > cpu）
         """
         # 如果显式要求 CPU，直接返回（唯一有意义的强制场景）
+        if os.environ.get('FORCE_CPU') == '1':
             return torch.device("cpu")
         # 自动选择最优设备
         load_description = "模型"
         # 环境变量配置
+        # FORCE_INT8=1: 启用 INT8 量化（适用于 CPU 和 CUDA，实验性，在某些情况下会降低性能）
+        # CPU_FORCE_BFLOAT16=1: 启用 bfloat16（仅适用于 CPU，需硬件加速支持，否则会降低性能）
+        force_int8 = os.environ.get('FORCE_INT8') == '1'
+        force_bfloat16 = os.environ.get('CPU_FORCE_BFLOAT16') == '1'
         # 检测是否为 AWQ 模型（自动检测）
         is_awq_model = self._is_awq_model(model_path)
                 use_int8 = True
                 device_map = "cpu"
                 load_description = "模型（INT8量化）"
+                print("⚠️  启用 INT8 量化（FORCE_INT8=1，实验性，在某些情况下会降低性能）")
             elif force_bfloat16:
                 dtype = torch.bfloat16
                 use_low_cpu_mem = True
+                print("⚠️  启用 bfloat16（CPU_FORCE_BFLOAT16=1，需硬件加速支持，否则会降低性能）")
             else:
                 # 默认: float32
             if force_int8:
                 use_int8 = True
                 load_description = "模型（INT8量化）"
+                print("⚠️  启用 INT8 量化（FORCE_INT8=1）")
             else:
                 dtype = torch.float16
                 print("🔧 dtype: float16")
             print(f"🔧 {self.device.type.upper()} 模式：自动设备分配")
             if force_int8:
+                print("⚠️  MPS 不支持 INT8 量化，已忽略 FORCE_INT8=1 环境变量")
             device_map = "auto"
             dtype = torch.float16
         device_name = DeviceManager.get_device_name(self.device)
         print(f"✓ {model_display_name} 模型已加载 ({device_name})")
+        # 初始化分析计数器（用于控制GPU内存统计打印频率）
+        self._analysis_count = 0
     def _load_model_with_int8_cuda(
         self,
             DeviceManager.clear_cache(self.device)
             gc.collect()
+            # 更新分析计数器
+            self._analysis_count += 1
+            # 打印分析任务完成后的内存统计（第1、11、21...次分析后打印）
+            if self.device.type == "cuda" and (self._analysis_count - 1) % 10 == 0:
                 device_idx = self.device.index if self.device.index is not None else 0
                 DeviceManager.print_cuda_memory_summary(device=device_idx)

backend/runtime_config.py CHANGED Viewed

@@ -116,7 +116,7 @@ def detect_platform(verbose: bool = True) -> str:
         平台 ID 字符串（如 'local_mps', 'cloud_cuda', 'cloud_cpu_16g', 'default_cpu_machine'）
     """
     # 1. 显式强制 CPU
-    if os.environ.get("FORCE_CPU"):
         print(f"🔧 强制 CPU 模式")
         return _detect_cpu_variant()

         平台 ID 字符串（如 'local_mps', 'cloud_cuda', 'cloud_cpu_16g', 'default_cpu_machine'）
     """
     # 1. 显式强制 CPU
+    if os.environ.get("FORCE_CPU") == "1":
         print(f"🔧 强制 CPU 模式")
         return _detect_cpu_variant()

client/src/content/home.en.html CHANGED Viewed

@@ -75,8 +75,9 @@
             AI-generated nonsense with no information content.</p>
         <p><strong>What LLM is currently used?</strong></p>
-        <p>Currently <strong>Qwen3-14B-Base</strong> is used, which gives pretty good results among the
-            models the author has tested.</p>
         <p><strong>Why does information content affect text quality?</strong></p>
         <p>Low information content means the LLM can easily predict it from context. If even a machine can predict it,

             AI-generated nonsense with no information content.</p>
         <p><strong>What LLM is currently used?</strong></p>
+        <p>Currently the open-source <strong>Qwen3-14B-Base</strong> is used, which gives pretty good results among the
+            models the author has tested. When lack of hardware credits, <strong>Qwen3-0.6B-Base</strong> is used
+            instead; it's smaller, faster, and performs slightly worse than Qwen3-14B-Base (about 30%).</p>
         <p><strong>Why does information content affect text quality?</strong></p>
         <p>Low information content means the LLM can easily predict it from context. If even a machine can predict it,

client/src/content/home.zh.html CHANGED Viewed

@@ -53,7 +53,7 @@
         </p>
         <p><strong>目前使用的是什么大模型？</strong></p>
-        <p>当前使用的是开源的 <strong>Qwen3-14B-Base</strong>，它是作者测试过的模型里结果挺不错的一个。</p>
         <p><strong>说到底，为什么信息量会影响文本的质量？</strong></p>
         <p>一个词的信息量低，意味着大模型能很容易从上文预测出来。既然机器都能预测出来，那它还能有多关键呢？反之，一个词的信息量高，意味着大模型很难从上文预测出来。（如果不是错误表达的话）那它就代表了作者想要表达，而机器不知道的关键信息。

         </p>
         <p><strong>目前使用的是什么大模型？</strong></p>
+        <p>当前使用的是开源的 <strong>Qwen3-14B-Base</strong>，它是作者测试过的模型里结果挺不错的一个。当硬件额度不足时，会用Qwen3-0.6B-Base模型，它体积小，速度快，效果比Qwen3-14B-Base稍差（30%左右）。</p>
         <p><strong>说到底，为什么信息量会影响文本的质量？</strong></p>
         <p>一个词的信息量低，意味着大模型能很容易从上文预测出来。既然机器都能预测出来，那它还能有多关键呢？反之，一个词的信息量高，意味着大模型很难从上文预测出来。（如果不是错误表达的话）那它就代表了作者想要表达，而机器不知道的关键信息。

client/src/index.html CHANGED Viewed

@@ -3,7 +3,9 @@
 <head>
     <meta charset="UTF-8">
-    <title>InfoRadar(信息雷达)</title>
     <meta name="viewport" content="width=device-width, initial-scale=1.0">
     <link rel="stylesheet" type="text/css" href="start.css">
 </head>

 <head>
     <meta charset="UTF-8">
+    <title>InfoRadar — Analyze Text Information Density</title>
+    <meta name="description"
+      content="InfoRadar visualizes token-level information density in text using LLMs, helping you quickly find key content and skip redundancy.">
     <meta name="viewport" content="width=device-width, initial-scale=1.0">
     <link rel="stylesheet" type="text/css" href="start.css">
 </head>

data/demo/public/CN/GPT-2 large unicorn text(中文翻译).json ADDED Viewed