---
# For reference on dataset card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/datasetcard.md?plain=1
# Doc / guide: https://huggingface.co/docs/hub/datasets-cards
{}
---

# Dataset Card for LLMcoder-GitHub-Python-Mix-Direct

Python target autocomplete suggestions in the format of conversations for OpenAI's fine-tuning.

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->



- **Curated by:** [More Information Needed]
- **Funded by [optional]:** [More Information Needed]
- **Shared by [optional]:** [More Information Needed]
- **Language(s) (NLP):** [More Information Needed]
- **License:** [More Information Needed]

### Dataset Sources [optional]

The data has been scraped from the following public GitHub repositories on 2023-11-15:

```
https://github.com/numpy/numpy
https://github.com/pandas-dev/pandas
https://github.com/matplotlib/matplotlib
https://github.com/scikit-learn/scikit-learn
https://github.com/python-pillow/Pillow
https://github.com/psaegert/pmtrendviz
https://github.com/psaegert/nli-nec
https://github.com/graphdeco-inria/gaussian-splatting
https://github.com/lllyasviel/ControlNet
https://github.com/maltfield/awesome-lemmy-instances
https://github.com/Aleph-Alpha/aleph-alpha-client
https://github.com/MaartenGr/BERTopic
https://github.com/MilesCranmer/PySR
https://github.com/AUTOMATIC1111/stable-diffusion-webui
https://github.com/microsoft/Codex-CLI
https://github.com/dropbox/hydra
https://github.com/HLearning/unet_keras
https://github.com/hmason/ml_class
https://github.com/django/django
https://github.com/encode/django-rest-framework
https://github.com/pallets/flask
https://github.com/postmanlabs/httpbin
https://github.com/jakevdp/PythonDataScienceHandbook
https://github.com/donnemartin/data-science-ipython-notebooks
https://github.com/tensorflow/tensorflow
```

## Uses

This dataset is intended to be used for fine-tuning of GPT-3.5-Turbo via OpenAI's fine-tuning API.

## Dataset Structure

`train_completions.jsonl` includes a list of 100 target conversations. Each conversation is structured as follows:

```json
[
    {"role": "system", "content": <system prompt>},
    {"role": "user", "content": <first half of the code>},
    {"role": "assistant", "content": <small target completion from ground truth>}
]
```

## Dataset Creation

### Curation Rationale

This dataset was created to fine-tune GPT-3.5-Turbo to provide more reliably formatted autocomplete suggestions for python code.

#### Data Collection and Processing

We scraped a mix of 25 GitHub repositories related to python and sampled 4 random python files weighted by their length.
The files were split in 2 parts at a uniformly random point in each file.
Next, if the input was longer than 10k tokens, we truncated the input file to 250 to 10000 tokens from the beginning.
The output was manually truncated to reasonably short code completions.

#### Who are the source data producers?

Paul Saegert

[More Information Needed]

#### Personal and Sensitive Information

This dataset may contain personal or sentitive information.

## Bias, Risks, and Limitations

The dataset contains only python code from the most popular, trending, or personal projects. It may be biased towards a particular style of code.

## Citation [optional]

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

[More Information Needed]