Research Papers 论文研究 6h ago Updated 1h ago 更新于 1小时前 50

SynthSentry: Detecting Synthetic Data Contamination in Language Model Training Data SynthSentry:检测语言模型训练数据中的合成数据污染

SynthSentry is a corpus-level, model-agnostic method for detecting synthetic data contamination in LLM training data without requiring access to the generating model, generation history, or synthetic labels The scoring mechanism combines three distributional divergence statistics: lexical diversity collapse, n-gram tail truncation, and perplexity variance across reference models A naive quantile-based threshold runs four times over the nominal false-positive budget; replacing it with covariance 提出SynthSentry,一种无需访问生成模型、生成历史或合成标签的语料库级合成数据污染检测方案 通过词汇多样性崩溃、n-gram尾部截断、跨参考模型困惑度方差三个统计量的分布差异计算污染分数 采用协方差收缩与bootstrap阈值替代朴素分位数,使假阳性率控制在名义预算附近(朴素方法超标4倍) 在小型开源生成器和指令微调模型污染的语料库上验证,对自然重复文本(法律、临床、代码)的假阳性可控 将污染检测定位为数据策展防御而非事后诊断,并开源评分工具包

68
Hot 热度
76
Quality 质量
72
Impact 影响力

Analysis 深度分析

TL;DR

  • SynthSentry is a corpus-level, model-agnostic method for detecting synthetic data contamination in LLM training data without requiring access to the generating model, generation history, or synthetic labels
  • The scoring mechanism combines three distributional divergence statistics: lexical diversity collapse, n-gram tail truncation, and perplexity variance across reference models
  • A naive quantile-based threshold runs four times over the nominal false-positive budget; replacing it with covariance shrinkage and a bootstrap threshold restores per-domain calibration
  • The toolkit is released as a data-curation defense for pre-training screening, though results are limited to small-scale, English-language, batch-mode corpus screening with single-generation contamination sources

Why It Matters

As synthetic data becomes increasingly prevalent in training corpora, detecting contamination before training is critical to preventing model collapse and preserving factual accuracy. SynthSentry provides a practical, model-agnostic screening tool that shifts the paradigm from post-hoc diagnosis to proactive data curation, which is essential for any organization building or fine-tuning language models on mixed provenance data.

Technical Details

  • Three-statistic scoring: The SynthSentry score measures distributional divergence across (1) lexical diversity collapse, (2) n-gram tail truncation, and (3) perplexity variance computed across multiple reference models
  • Model-agnostic design: Requires no knowledge of the generating model, no generation history, and no synthetic labels, making it applicable to corpora of unknown provenance
  • Threshold calibration: A naive quantile-based approach was found to produce false positives at four times the nominal budget; the authors replace this with covariance shrinkage and a bootstrap threshold to achieve proper per-domain calibration
  • Evaluation protocol: Tested on corpora contaminated by small open-weight generators and an instruction-tuned open-weight model using a leave-one-generator-out protocol, with domain-stratified analysis on naturally repetitive human text (legal, clinical, source code)
  • Limitations: Results are small-scale, English-only, batch-mode; contamination sources are single-generation or hand-authored rather than recursively generated, so findings generalize to synthetic contamination broadly but not specifically to recursion depth

Industry Insight

  • Organizations ingesting third-party or web-scraped data should integrate pre-training contamination screening into their data curation pipelines; SynthSentry offers a ready-to-use toolkit for this purpose
  • The finding that naive statistical thresholds dramatically overfit false positives underscores the importance of careful calibration—especially when screening domain-specific corpora like legal or clinical text where natural repetition is expected
  • While the paper shows no contamination-driven accuracy deficit at their scale, the open question of whether pruning recovers lost performance means practitioners should treat SynthSentry as a risk-ranking tool rather than a definitive quality guarantee, and monitor over-pruning risks once removal exceeds true contamination fractions

TL;DR

  • 提出SynthSentry,一种无需访问生成模型、生成历史或合成标签的语料库级合成数据污染检测方案
  • 通过词汇多样性崩溃、n-gram尾部截断、跨参考模型困惑度方差三个统计量的分布差异计算污染分数
  • 采用协方差收缩与bootstrap阈值替代朴素分位数,使假阳性率控制在名义预算附近(朴素方法超标4倍)
  • 在小型开源生成器和指令微调模型污染的语料库上验证,对自然重复文本(法律、临床、代码)的假阳性可控
  • 将污染检测定位为数据策展防御而非事后诊断,并开源评分工具包

为什么值得看

随着合成数据在LLM训练中的广泛使用,如何识别和过滤污染语料成为关键工程问题。本文提供了可操作的预训练筛查方案,填补了"训练前防御"而非"训练后诊断"的方法空白,对数据质量管控具有实用价值。

技术解析

  • 核心指标设计:污染分数基于三个统计量的分布散度——词汇多样性坍缩(lexical diversity collapse)、n-gram尾部截断(n-gram tail truncation)、跨参考模型的困惑度方差(perplexity variance across reference models),三者共同捕捉合成文本的统计特征偏移。
  • 校准优化:原始朴素分位数方法假阳性超标4倍,引入协方差收缩(covariance shrinkage)和bootstrap阈值后,各域校准接近名义假阳性预算,显著改善误判率。
  • 评估协议:采用leave-one-generator-out协议,在小型开源生成器和指令微调模型污染的语料库上测试;同时通过domain-stratified研究评估法律、临床、源代码等自然重复文本的假阳性。
  • 局限性:实验规模较小,仅覆盖英语语料库的批量模式筛选;污染来源为单次生成或手工编写,未涉及递归生成场景,结论不直接适用于递归深度问题。

行业启示

  • 数据策展应前置为训练流程的防御环节,而非依赖事后诊断;建立标准化的合成数据筛查工具链将成为大模型训练的基础设施。
  • 合成数据污染检测需兼顾"检出率"与"假阳性控制",尤其在法律、医疗、代码等专业领域,过度过滤可能导致有价值数据的丢失。
  • 当前方法对单次生成污染有效,但递归生成场景下的模型崩溃风险仍需独立研究;业界需关注递归训练链路的长期数据质量衰减问题。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Training 训练 Dataset 数据集 Research 科学研究 Security 安全