README.md

---
license: other
dataset_info:
  features:
  - name: path
    dtype: string
  - name: owner
    dtype: string
  - name: repo_id
    dtype: int64
  - name: is_fork
    dtype: bool
  - name: languages_distribution
    dtype: string
  - name: content
    dtype: string
  - name: issues
    dtype: float64
  - name: main_language
    dtype: string
  - name: forks
    dtype: int64
  - name: stars
    dtype: int64
  - name: commit_sha
    dtype: string
  - name: size
    dtype: int64
  - name: name
    dtype: string
  - name: license
    dtype: string
  splits:
  - name: train
    num_bytes: 75063445
    num_examples: 25000
  download_size: 29298620
  dataset_size: 75063445
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---


# Dataset Summary
The dataset contains 25,000 Kotlin code samples selected from the [KStack](https://huggingface.co/datasets/JetBrains/KStack) dataset. The selection is performed based on the value of the code for learning algorithmic concepts in Kotlin. In total, the dataset contains about 23M [CodeLlama-7b](https://huggingface.co/codellama/CodeLlama-7b-hf) tokens (vocab size 32,016).

# Dataset Collection
The filtering from [KStack](https://huggingface.co/datasets/JetBrains/KStack) is performed using zero-shot quality estimation based on [Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2). The model is prompted to determine which of two files has higher "educational value for learning algorithms in Kotlin". The results of the comparisons are averaged and used to train a binary classifier based on [CodeT5p-220m](https://huggingface.co/Salesforce/codet5p-220m). The binary classifier is then applied to the entire KStack to obtain scores for each sample in the dataset. The log-probability of the classifier prediction is used as a criterion of the selection.

# Opt-out
If you want your data to be removed from dataset, or have any other questions, please reach out to Sergey Titov: <[email protected]>