README.md

metadata

language:
  - en
license: other
tags:
  - code
  - seeds
  - oss-instruct
  - dataset-curation

oss-code-seeds

A combined dataset of open-source code snippets (seeds) curated from multiple high-quality sources. The intended use is to serve as seeds for generating coding instruction-response pairs via the OSS-Instruct approach — where a model is prompted with a real code snippet to produce a grounded, diverse coding problem and its solution.

Columns

• seed: A raw code snippet from open-source software (OSS), serving as the seed
• source: The original HuggingFace dataset the seed was taken from

Sources

Dataset	Description
ise-uiuc/Magicoder-OSS-Instruct-75K	Seeds used to generate the Magicoder dataset via GPT-3.5, multi-language GitHub snippets
bigcode/self-oss-instruct-sc2-concepts	Filtered Python functions from The Stack V1 used in the SelfCodeAlign pipeline

Intended Use

Feed the seed column into your own model API to generate coding problems and solutions, effectively replicating or improving upon the OSS-Instruct pipeline with your own model.

from datasets import load_dataset

ds = load_dataset("PursuitOfDataScience/oss-code-seeds", split="train")

for row in ds:
    seed = row["seed"]
    source = row["source"]
    # prompt your model with seed to generate a problem + solution