metadata
license: mit
language:
- en
task_categories:
- text-generation
pretty_name: COLLIE
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: test
path: test.parquet
COLLIE: Systematic Construction of Constrained Text Generation Tasks
- Paper: https://arxiv.org/abs/2307.08689
- Code / original data: https://github.com/princeton-nlp/Collie (MIT)
- Source file:
https://collie-benchmark.github.io/data/all_data.dill
Contents
1,306 constrained-generation instances (one test split) covering 13 constraint
structures across word, sentence, paragraph and passage level. Text is drawn from
Wikipedia, Project Gutenberg, CC-News and an English word list.
| column | type | description |
|---|---|---|
source |
str | wiki, guten, ccnews or english |
constraint_id |
str | one of c01–c05, c06a, c07–c12, c14 (paper Table 1) |
constraint |
str | human-readable rendering of the constraint structure |
prompt |
str | natural-language instruction given to the model |
example |
str | a reference text from the corpus that satisfies the constraint |
targets |
str (JSON) | instance-specific constraint values (int, list of str, …) |
metadata |
str (JSON) | source document metadata (title, index, author) |
Citation
@misc{yao2023collie,
title={{COLLIE}: Systematic Construction of Constrained Text Generation Tasks},
author={Shunyu Yao and Howard Chen and Austin W. Hanjie and Runzhe Yang and Karthik Narasimhan},
year={2023},
eprint={2307.08689},
archivePrefix={arXiv},
primaryClass={cs.CL}
}