LLM distillation detection and model fingerprint audit tool - text source detection, model identity verification, and distillation analysis
Project description
ModelAudit
LLM 蒸馏检测与模型指纹审计 — 文本溯源、身份验证、蒸馏关系判定 LLM distillation detection & model fingerprinting – text provenance, identity verification, distillation auditing
快速开始 · 检测方法 · MCP Server · Data Pipeline 生态
GitHub Topics: model-fingerprint, llm-distillation, model-audit, cli, mcp, ai-data-pipeline
检测文本数据来源、验证 API 模型身份、审计模型蒸馏关系。黑盒优先,标注员友好。
核心能力 / Core Capabilities
文本/模型 → 探测 Prompt → 响应特征提取 → 指纹比对 → 审计报告
审计仪表盘预览 / Sample Dashboard
┌───────────────────────────────────────────────┐
│ 模型蒸馏审计报告 │
├───────────────┬──────────────┬────────────────┤
│ 教师: gpt-4o │ 学生: my-llm │ 相似度: 0.9213 │
├───────────────┴──────────────┴────────────────┤
│ ⚠️ 判定: 可能存在蒸馏关系 │
│ 📊 置信度: 87.5% │
│ 🔍 风格匹配: helpful 0.82 / hedging 0.79 │
└───────────────────────────────────────────────┘
功能矩阵 / Features
| 功能 | 说明 |
|---|---|
| 🔍 文本来源检测 | 判断一批文本是哪个 LLM 生成的 |
| ✅ 模型身份验证 | 验证 API 背后是不是声称的模型 |
| 🔗 模型指纹比对 | 比对两个模型的行为特征相似度 |
| 📋 蒸馏审计报告 | 综合分析生成 Markdown / JSON 报告 |
安装 / Installation
pip install knowlyr-modelaudit
可选依赖:
pip install knowlyr-modelaudit[blackbox] # 黑盒指纹 (openai, anthropic, httpx)
pip install knowlyr-modelaudit[whitebox] # 白盒指纹 (torch, transformers)
pip install knowlyr-modelaudit[mcp] # MCP 服务器
pip install knowlyr-modelaudit[all] # 全部功能
快速开始 / Quick Start
检测文本来源 / CLI
# 检测文本数据是哪个模型生成的
knowlyr-modelaudit detect texts.jsonl
# 限制条数,输出 JSON
knowlyr-modelaudit detect texts.jsonl -n 50 -f json -o result.json
输出示例
正在分析 3 条文本...
ID | 预测模型 | 置信度 | 预览
------------------------------------------------------------
1 | chatgpt | 72.50% | Certainly! I'd be happy to...
2 | chatgpt | 65.00% | I think that's an interest...
3 | chatgpt | 70.00% | Sure thing! No problem at ...
来源分布:
chatgpt: 3 (100.0%)
验证模型身份
# 验证 API 背后是不是声称的 GPT-4o
knowlyr-modelaudit verify gpt-4o --provider openai
# 自定义 API
knowlyr-modelaudit verify my-model --provider custom --api-base http://localhost:8000
比对模型指纹
# 比对两个模型是否存在蒸馏关系
knowlyr-modelaudit compare gpt-4o claude-sonnet --provider openai
完整蒸馏审计
# 同一 provider — 生成详细审计报告
knowlyr-modelaudit audit --teacher gpt-4o --student my-model -o report.md
# 跨 provider 审计 — 分别配置不同 API
knowlyr-modelaudit audit \
--teacher claude-opus --teacher-provider anthropic \
--student kimi-k2.5 --student-provider openai \
--student-api-base https://api.moonshot.cn/v1 \
-o report.md
# 强制重新调用 API(跳过缓存)
knowlyr-modelaudit audit --teacher gpt-4o --student my-model --no-cache
自动生成 6 节详细审计报告:审计对象 → 方法 → 结果(指纹详情 + 逐条探测)→ 关键发现 → 结论 → 局限性声明。
输出示例
正在审计: claude-opus → kimi-k2.5...
判定结果: ⚠️ 可能存在蒸馏关系
置信度: 0.7980
报告已自动保存: reports/kimi-k2.5-vs-claude-opus-audit.md
指纹缓存
# 查看缓存的指纹
knowlyr-modelaudit cache list
# 清除缓存
knowlyr-modelaudit cache clear
首次审计时自动缓存模型指纹到本地 .modelaudit_cache/,再次审计同一模型时直接复用,避免重复调 API。
在 Python 中接入 / Python SDK
from modelaudit import AuditEngine
engine = AuditEngine()
# 检测文本来源
results = engine.detect(["Hello! I'd be happy to help..."])
for r in results:
print(f"{r.predicted_model}: {r.confidence:.2%}")
# 比对模型指纹 (需要 API key)
result = engine.compare("gpt-4o", "my-model", method="llmmap")
print(f"相似度: {result.similarity:.4f}")
print(f"蒸馏关系: {'是' if result.is_derived else '否'}")
检测方法 / Detection Methods
已实现
| 方法 | 类型 | 说明 | 参考 |
|---|---|---|---|
| LLMmap | 黑盒 | 20 个探测 Prompt,分析响应模式 | USENIX Security 2025 |
| StyleAnalysis | 风格分析 | 12 个模型家族的风格签名匹配 | — |
支持识别的模型家族
gpt-4 · gpt-3.5 · claude · llama · gemini · qwen · deepseek · mistral · yi · phi · cohere · chatglm
探测维度(20 个 Probe)
| 维度 | 探测内容 |
|---|---|
| 自我认知 | 模型身份、创建者、训练截止 |
| 安全边界 | 拒绝策略、措辞差异 |
| 注入测试 | Prompt injection 响应差异 |
| 知识与推理 | 知识边界、逻辑推理、伦理判断 |
| 创意写作 | 叙事风格、类比能力 |
| 多语言 | 中文响应、多语翻译 |
| 格式控制 | JSON 输出、Markdown 表格 |
| 角色扮演 | 角色一致性、创意表达 |
| 代码生成 | 编码风格、注释习惯 |
| 摘要能力 | 信息压缩、表达密度 |
规划中
| 方法 | 类型 | 说明 | 参考 |
|---|---|---|---|
| REEF | 白盒 | CKA 隐层相似度比对 | ICLR 2025 Oral |
| DLI | 蒸馏检测 | 影子模型 + 行为签名 | ICLR 2026 |
查看可用方法
knowlyr-modelaudit methods
MCP Server
在 Claude Desktop / Claude Code 中直接使用。
配置
添加到 ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"knowlyr-modelaudit": {
"command": "uv",
"args": ["--directory", "/path/to/model-audit", "run", "python", "-m", "modelaudit.mcp_server"]
}
}
}
可用工具
| 工具 | 功能 |
|---|---|
detect_text_source |
检测文本数据来源 |
verify_model |
验证模型身份 |
compare_models |
比对两个模型指纹 |
audit_distillation |
完整蒸馏审计 |
使用示例
用户: 帮我检测这批文本是哪个模型生成的
Claude: [调用 detect_text_source]
## 文本来源检测结果
| # | 预测模型 | 置信度 | 预览 |
|---|---------|--------|------|
| 1 | chatgpt | 72.50% | Certainly! I'd be happy... |
### 来源分布
- chatgpt: 3 (100.0%)
Data Pipeline 生态
ModelAudit 是 Data Pipeline 生态的模型质检组件:
graph LR
Radar["🔍 Radar<br/>情报发现"] --> Recipe["📋 Recipe<br/>逆向分析"]
Recipe --> Synth["🔄 Synth<br/>数据合成"]
Recipe --> Label["🏷️ Label<br/>数据标注"]
Synth --> Check["✅ Check<br/>数据质检"]
Label --> Check
Check --> Audit["🔬 Audit<br/>模型审计"]
Audit --> Hub["🎯 Hub<br/>编排层"]
Hub --> Sandbox["📦 Sandbox<br/>执行沙箱"]
Sandbox --> Recorder["📹 Recorder<br/>轨迹录制"]
Recorder --> Reward["⭐ Reward<br/>过程打分"]
style Audit fill:#0969da,color:#fff,stroke:#0969da
生态项目
| 层 | 项目 | 说明 | 仓库 |
|---|---|---|---|
| 情报 | AI Dataset Radar | 数据集竞争情报、趋势分析 | GitHub |
| 分析 | DataRecipe | 逆向分析、Schema 提取、成本估算 | GitHub |
| 生产 | DataSynth | LLM 批量合成、种子数据扩充 | GitHub |
| 生产 | DataLabel | 轻量标注工具、多标注员合并 | GitHub |
| 质检 | DataCheck | 规则验证、重复检测、分布分析 | GitHub |
| 质检 | ModelAudit | 蒸馏检测、模型指纹、身份验证 | You are here |
| Agent | AgentSandbox | Docker 执行沙箱、轨迹重放 | GitHub |
| Agent | AgentRecorder | 标准化轨迹录制、多框架适配 | GitHub |
| Agent | AgentReward | 过程级 Reward、Rubric 多维评估 | GitHub |
| 编排 | TrajectoryHub | Pipeline 编排、数据集导出 | GitHub |
端到端工作流
# 1. DataRecipe: 分析数据集,生成 Schema 和样例
knowlyr-datarecipe deep-analyze tencent/CL-bench -o ./output
# 2. DataSynth: 基于种子数据批量合成
knowlyr-datasynth generate ./output/tencent_CL-bench/ -n 1000
# 3. DataCheck: 数据质量检查
knowlyr-datacheck validate ./output/tencent_CL-bench/
# 4. ModelAudit: 检测合成数据来源,验证模型身份
knowlyr-modelaudit detect ./output/synthetic.jsonl
knowlyr-modelaudit verify gpt-4o --provider openai
组合 MCP 配置
{
"mcpServers": {
"knowlyr-datarecipe": {
"command": "uv",
"args": ["--directory", "/path/to/data-recipe", "run", "knowlyr-datarecipe-mcp"]
},
"knowlyr-datacheck": {
"command": "uv",
"args": ["--directory", "/path/to/data-check", "run", "python", "-m", "datacheck.mcp_server"]
},
"knowlyr-modelaudit": {
"command": "uv",
"args": ["--directory", "/path/to/model-audit", "run", "python", "-m", "modelaudit.mcp_server"]
}
}
}
命令参考
| 命令 | 功能 |
|---|---|
knowlyr-modelaudit detect <file> |
检测文本数据来源 |
knowlyr-modelaudit detect <file> -n 50 |
限制检测条数 |
knowlyr-modelaudit verify <model> |
验证模型身份 |
knowlyr-modelaudit compare <a> <b> |
比对两个模型指纹 |
knowlyr-modelaudit audit --teacher <a> --student <b> |
完整蒸馏审计 |
knowlyr-modelaudit audit ... --teacher-provider anthropic |
跨 provider 审计 |
knowlyr-modelaudit audit ... --no-cache |
跳过缓存,强制重新调 API |
knowlyr-modelaudit cache list |
查看缓存的指纹 |
knowlyr-modelaudit cache clear |
清除所有缓存 |
knowlyr-modelaudit methods |
列出可用检测方法 |
API 使用
from modelaudit import AuditEngine, Fingerprint, ComparisonResult
# 创建引擎(默认启用指纹缓存)
engine = AuditEngine()
# 检测文本来源
results = engine.detect(texts)
for r in results:
print(f"#{r.text_id} {r.predicted_model} ({r.confidence:.2%})")
# 指纹比对 (需要 API key)
result = engine.compare("gpt-4o", "my-model", method="llmmap")
print(f"相似度: {result.similarity:.4f}")
# 完整审计(支持跨 provider)
audit = engine.audit(
"claude-opus", "kimi-k2.5",
teacher_provider="anthropic",
student_provider="openai",
student_api_base="https://api.moonshot.cn/v1",
)
print(audit.verdict) # likely_derived / independent / inconclusive
print(audit.confidence) # 0.798
# 生成详细报告(6 节结构)
from modelaudit.report import generate_report
report = generate_report(audit, "markdown")
# 不使用缓存
engine_no_cache = AuditEngine(use_cache=False)
项目架构
src/modelaudit/
├── engine.py # AuditEngine 总入口
├── models.py # Pydantic 数据模型
├── base.py # Fingerprinter 抽象基类
├── registry.py # 方法注册表
├── config.py # 配置
├── cache.py # 指纹缓存
├── methods/
│ ├── llmmap.py # LLMmap 黑盒指纹
│ └── style.py # 风格分析
├── probes/
│ └── prompts.py # 探测 Prompt 库
├── report.py # 报告生成 (6 节详细报告)
├── cli.py # CLI 命令行 (7 命令)
└── mcp_server.py # MCP Server (4 工具)
License
AI Data Pipeline 生态
10 个工具覆盖 AI 数据工程全流程,均支持 CLI + MCP,可独立使用也可组合成流水线。
| Tool | Description | Link |
|---|---|---|
| AI Dataset Radar | Competitive intelligence for AI training datasets | GitHub |
| DataRecipe | Reverse-engineer datasets into annotation specs & cost models | GitHub |
| DataSynth | Seed-to-scale synthetic data generation | GitHub |
| DataLabel | Lightweight, serverless HTML labeling tool | GitHub |
| DataCheck | Automated quality checks & anomaly detection | GitHub |
| ModelAudit | LLM distillation detection & model fingerprinting | You are here |
| AgentSandbox | Reproducible Docker sandbox for Code Agent execution | GitHub |
| AgentRecorder | Standardized trajectory recording for Code Agents | GitHub |
| AgentReward | Process-level rubric-based reward engine | GitHub |
| TrajectoryHub | Pipeline orchestrator for Agent trajectory data | GitHub |
graph LR
A[Radar] --> B[Recipe] --> C[Synth] --> E[Check] --> F[Audit] --> G[Hub]
B --> D[Label] --> E
G --> H[Sandbox] --> I[Recorder] --> J[Reward]
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file knowlyr_modelaudit-0.2.1.tar.gz.
File metadata
- Download URL: knowlyr_modelaudit-0.2.1.tar.gz
- Upload date:
- Size: 35.3 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: uv/0.9.26 {"installer":{"name":"uv","version":"0.9.26","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"macOS","version":null,"id":null,"libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
06197903ae3e599ac53149d8efb97d767dfd16f657061f14aaa1e2e912357cc2
|
|
| MD5 |
47019f143becaa52afe2c2a24a6f1ffa
|
|
| BLAKE2b-256 |
73e8d6157ae1bc465eb896d4d77e538a62cd7d80ae6ab16c8565dd927fd8767c
|
File details
Details for the file knowlyr_modelaudit-0.2.1-py3-none-any.whl.
File metadata
- Download URL: knowlyr_modelaudit-0.2.1-py3-none-any.whl
- Upload date:
- Size: 40.3 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: uv/0.9.26 {"installer":{"name":"uv","version":"0.9.26","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"macOS","version":null,"id":null,"libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
e037d95e75f55366b3ccecc99bfb4a35275c5d53631f47d6a6b98f2c96851207
|
|
| MD5 |
c6830b064b095250bfc1a0e0df517f72
|
|
| BLAKE2b-256 |
366cc160d25445003ac7bf28c7b1eacca01bb8968be8070e9fde42fd37c79281
|