Python Codes 25K: 24,813 Cleaned Instructional Tasks for Python

A set of 24,813 cleaned Python instructional tasks published by flytech in May 2024. It provides text-to-code and code-to-text pairs formatted for instruction tuning, featuring an average of 508 tokens per entry.

Use Cases

• Fine-tuning large language models for Python code generation using the 'instruction' and 'output' columns
• Training code summarization models by mapping 'output' code snippets back to 'instruction' descriptions
• Developing instruction-following assistants using the pre-formatted 'text' field

Strengths

• 24,813 total entries
• High instruction diversity with 24,580 unique instructions
• MIT license allows for broad reuse
• Average length of 508 tokens per example provides sufficient context for training

Limitations

• Low input diversity with only 3,666 unique inputs across 24,813 records
• Restricted exclusively to the Python programming language
• Potential for synthetic data artifacts common in instruction-following datasets

Provenance

Source

flytech via Hugging Face

Freshness

Last updated May 2024; no scheduled updates indicated.