metadata
dataset_info:
features:
- name: joke
dtype: string
- name: upvotes
dtype: int64
- name: downvotes
dtype: int64
splits:
- name: train
num_bytes: 3074127
num_examples: 9020
download_size: 2061760
dataset_size: 3074127
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: mit
task_categories:
- text-generation
language:
- pl
pretty_name: Dowcipy jaja
tags:
- art
- not-for-all-audiences
size_categories:
- 1K<n<10K
Dataset consisting of polish jokes
Warning: Jokes were not curated, some may be offensive, stupid or simply not funny. It's highly recommended to filter jokes before training, e.g., based on downvotes
This dataset consists of all (9k) jokes dumped from jeja.pl on 2024-02-14. Jokes are submitted by the community. Besides the funny text itself, I included upvotes and downvotes. You can use them for filtering.
Default sorting is based on a combination of downvotes and upvotes.
If used for training LLMs, it's recommended to use a tokenizer that supports line breaks, as these are often important for readability of the jokes.