metadata

license: mit
language:
  - en
tags:
  - code

Filtered StarCoder Dataset Mini

Dataset Description

This dataset contains filtered and processed code samples from 10 popular programming languages: C, C++, C#, Go, Java, JavaScript, Python, Ruby, Scala, and TypeScript. The dataset was created by filtering source code based on quality metrics, removing outliers, and standardizing the format for machine learning and code analysis applications.

Key Features

Cleaned and Filtered Code: Samples have been processed to remove outliers in terms of line length and code size
Quality Metrics: Each sample includes metadata about average line length and line count
Multi-language Support: 10 programming languages represented in separate subsets
Consistent Format: All samples follow the same Parquet structure for easy processing

Dataset Size

The complete dataset is approximately 12GB in size. Individual language files vary in size, with the largest being C++ (2GB) and the smallest being Scala (665MB).

Dataset Statistics

Language	Sample Count	Avg. Line Length	Avg. Line Count
C	1,752,078	28.07	61.85
C++	1,769,333	28.16	87.99
C#	1,763,508	29.53	44.29
Go	1,751,120	25.18	68.26
Java	1,779,659	30.84	54.35
JavaScript	1,718,133	27.68	44.07
Python	1,764,099	32.68	54.66
Ruby	1,756,771	27.35	27.34
Scala	952,890	35.30	44.38
TypeScript	1,738,885	29.17	36.84

Dataset Structure

The dataset is organized with separate Parquet files for each programming language:

c.parquet - C language samples
cpp.parquet - C++ language samples
c-sharp.parquet - C# language samples
go.parquet - Go language samples
java.parquet - Java language samples
javascript.parquet - JavaScript language samples
python.parquet - Python language samples
ruby.parquet - Ruby language samples
scala.parquet - Scala language samples
typescript.parquet - TypeScript language samples

Within each file, data is stored with the following schema:

- language: string (the programming language of the code sample)
- code: string (the complete code content)
- avg_line_length: float (average character count per line)
- line_count: integer (total number of lines in the code)

Each sample is stored as a row in the Parquet file with these four columns.

How to Access the Dataset

Using the Hugging Face `datasets` Library

This dataset is hosted on the Hugging Face Hub and can be easily accessed using the datasets library.

Install the Required Library

pip install datasets

Import Library

from datasets import load_dataset

Load the Entire Dataset

dataset = load_dataset(
    "jugalgajjar/Filtered-StarCoder-Dataset-Mini"
)

Load a Specific Language

dataset = load_dataset(
    "jugalgajjar/Filtered-StarCoder-Dataset-Mini",
    data_files="scala.parquet"
)

Stream Data

dataset = load_dataset(
    "jugalgajjar/Filtered-StarCoder-Dataset-Mini",
    data_files="scala.parquet",
    streaming=True
)

Access Data Content (After Downloading)

try:
    for example in dataset["train"].take(5):
        print(example)
        print("-"*25)
except Exception as e:
    print(f"An error occurred: {e}")

Manual Download

You can also manually download specific language files from the Hugging Face repository page:

Visit https://huggingface.co/datasets/jugalgajjar/Filtered-StarCoder-Dataset-Mini
Navigate to the "Files" tab
Click on the language file you want to download (e.g., python.parquet)
Use the download button to save the file locally

Dataset Creation

This dataset was created through the following process:

Original code samples were collected from the StarCoder dataset (URL)
Statistical analysis was performed to identify quality metrics
Outliers were removed using IQR (Interquartile Range) method
Samples were filtered to remove excessively long or short code examples
Data was normalized and standardized across languages
Metadata (average line length and line count) was calculated for each sample
Final data was serialized in the efficient Parquet format for optimal storage and access speed

The processing pipeline included steps to:

Remove code samples with abnormal line lengths (potential formatting issues)
Filter out extremely long files (exceeding the 90th percentile)
Ensure consistent formatting and structure
Generate useful metadata for each example

Citation

If you use this dataset in your research or project, please cite it as follows:

@misc{fscdmini2025,
  author = {Jugal Gajjar, Kamalasankari Subramaniakuppusamy, Kaustik Ranaware},
  title = {Filtered CodeStar Dataset Mini},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/jugalgajjar/Filtered-StarCoder-Dataset-Mini}}
}

License

This dataset is released under the MIT License. See the LICENSE file for more details.