---
dataset_info:
  features:
  - name: code
    dtype: string
  - name: repo_path
    dtype: string
  - name: parsed_code
    dtype: string
  - name: quality_prob
    dtype: float64
  - name: learning_prob
    dtype: float64
  splits:
  - name: train
    num_bytes: 852705076967
    num_examples: 65509810
  download_size: 0
  dataset_size: 852705076967
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# Dataset Card for "starcoder_labeled"

[Starcoder data](https://huggingface.co/datasets/bigcode/starcoderdata), with several popular languages selected, short sequences filtered out, then labeled based on learning quality (educational value) and code quality.

A good heuristic is to take anything with `>.5` code quality and `>.3` learning quality.  But you may want to vary the thresholds by language, depending on your target task.