metadata
dataset_info:
features:
- name: conversations
list:
- name: content
dtype: string
- name: role
dtype: string
splits:
- name: train
num_bytes: 298946
num_examples: 1414
download_size: 295950
dataset_size: 298946
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: mit
task_categories:
- question-answering
language:
- en
tags:
- patterns
pretty_name: Puzzles and Patterns
size_categories:
- 1K<n<10K
Puzzles and Patterns 1k
This is a SUPER simple dataset containing ~1400 short puzzles, mostly based around numbers. The intention of this dataset is to teach reasoning LLMs to recognize patterns as fast and effectively as possible.
Why no reasoning content?
I have decided to exclude reasoning from this dataset entirely, this dataset should be used in training base on answer, which encourages exploration by the model, rather than purely mimicking a style.