CodeR-Pile: 1M-10M Synthetic Records for Code Embedding Models

CodeR-Pile contains between 1 million and 10 million synthetic records designed for training generalist code embedding models. Released by nebula2025 in 2025, the data supports code retrieval tasks through massive synthesis of text and code pairs in English and Chinese.

Use Cases

• Training embedding models for code retrieval using synthetic text-code pairs
• Developing bilingual semantic search tools for English and Chinese code documentation
• Benchmarking code-to-code similarity tasks using synthesized programming examples

Strengths

• Large scale with 1M to 10M records
• Bilingual support for English and Chinese documentation
• Methodology documented in Arxiv paper 2505.12697

Limitations

• Synthetic nature may introduce artifacts not found in human-written code or queries
• Restrictive CC BY-NC-SA 4.0 license prevents commercial application

Provenance

Source

nebula2025, Arxiv 2505.12697

Collection Method

synthetic

Freshness

Last updated October 2025.