data-check
Free数据质检工具 — 自动化质量检查、异常检测、分布分析，保障数据集质量。
FreeFree tier
About data-check
DataCheck is an open-source composable rule engine for multi-dimensional LLM training data quality validation. It provides 9 built-in rules covering completeness, validity, privacy, and consistency, along with IQR/Z-score statistical anomaly detection and an auto-fix pipeline for deduplication, whitespace stripping, and PII redaction. The system supports YAML custom rules, 4 preset rule sets (default, sft, preference, llm), LLM-assisted quality evaluation, and report diff to track quality changes over time. It exposes 11 MCP tools for AI IDE integration and outputs structured quality reports in Markdown, JSON, and HTML formats.
Key Features
9 built-in quality rules: required_fields, format, length_bounds, PII detection, garbled text, near-duplicate via n-gram Jaccard, language consistency
IQR and Z-score statistical anomaly detection for numerical and length features
Auto-fix pipeline: dedup, strip whitespace, PII redaction
Report diff for tracking quality changes over time
LLM-assisted quality evaluation for instruction clarity and response relevance
11 MCP tools for AI IDE integration
YAML custom rules and 4 preset rule sets (default, sft, preference, llm)
Output reports in Markdown, JSON, and HTML formats
Pros & Cons
Pros
- Open-source and free to use
- Composable rule engine with 9 built-in rules and custom YAML rules
- Statistical anomaly detection (IQR/Z-score) without external dependencies
- End-to-end validation, auto-fix, and report diff pipeline
- MCP server enables integration with AI IDEs like GitHub Copilot
Cons
- Focused specifically on LLM training data, not a general data cleaning tool
- Requires Python environment and setup
- Anomaly detection is designed for numerical/length features
- May need configuration for domain-specific quality rules
Best For
Validating LLM training data quality before entering the training pipelineDetecting PII, garbled text, and near-duplicates in datasetsEnsuring data completeness, uniqueness, validity, and complianceAutomated data cleaning and quality assurance pipeline with report diffIntegration with AI coding assistants via MCP for real-time quality checks
FAQ
What quality dimensions does DataCheck cover?
DataCheck covers four dimensions: Completeness, Uniqueness, Validity, and Compliance.
What anomaly detection methods are supported?
DataCheck supports IQR (interquartile range) and Z-score methods for detecting numerical and length anomalies.
Can I define custom quality rules?
Yes, you can define custom rules using YAML configuration files.
Does DataCheck support automatic fixing of issues?
Yes, it includes an auto-fix pipeline for deduplication, whitespace stripping, and PII redaction.
How can I integrate DataCheck with my IDE?
It exposes 11 MCP tools that can be used with AI IDE plugins for real-time quality validation.