Mixtral Magicoder: Source Code Samples Across Multiple Programming Languages

A collection of source code samples from multiple programming languages, created by author mesolitica and last updated on June 1, 2025. The dataset was generated using the 'Magicoder: Source Code Is All You Need' template, targeting at least 10,000 rows per language. It includes samples for languages such as C++, C#, CUDA, and Dockerfile, sourced from the deduplicated version of The Stack dataset.

Use Cases

• Training code generation models based on multi-language source code samples.
• Fine-tuning models for specific programming languages like C++ or C#.
• Benchmarking model performance on code from diverse domains, including CUDA and Dockerfiles.
• Studying programming language patterns and syntax across different codebases.

Strengths

• Targets a minimum of 10,000 rows per programming language, as stated for C++ (10,747 rows) and C# (10,193 rows).
• Includes source code from multiple distinct languages and domains, such as CUDA and Dockerfile.
• Built from a known source, the deduplicated 'The Stack' dataset, which suggests a foundation of cleaned code.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality and sampling methodology require manual inspection.

Provenance

Source

Samples drawn from the deduplicated 'The Stack' dataset on Hugging Face.

Collection Method

Generated using the 'Magicoder: Source Code Is All You Need' template, targeting a minimum row count per language.

Time Range

null

Freshness

Last updated 2025-06-01 16:36:24.

Geography

null