COLLIE: Systematic Construction of Constrained Text Generation Tasks

Authors: Shunyu Yao, Howard Chen, Austin W. Hanjie, Runzhe Yang, Karthik R Narasimhan

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform systematic experiments across five state-of-the-art instruction-tuned language models and analyze their performances to reveal shortcomings.
Researcher Affiliation Academia Department of Computer Science, Princeton University {shunyuy, hc22, hjwang, runzhey, karthikn}@princeton.edu
Pseudocode Yes Figure 2: Example COLLIE code for a simple number of words constraint.
Open Source Code Yes Project site with code and data: https://collie-benchmark.github.io.
Open Datasets Yes We extract constraint targets from three distinct data sources: Wikipedia (Wiki) (Foundation, 2022), Common Crawl News (CC-News) (Hamborg et al., 2017), and the Project Gutenberg Corpus (Guten) (Brooke et al., 2015).
Dataset Splits No The paper evaluates pre-trained language models in a zero-shot setting on the COLLIE-v1 dataset. Therefore, it does not describe train/validation dataset splits as part of its experimental setup for training models.
Hardware Specification No The paper states 'All experiments were run in July, 2023.' but does not provide specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python version, library versions) used for conducting the experiments.
Experiment Setup Yes Our main experiments in this paper focus on a zero-shot prompting setup... By default, we use a sampling temperature of 0.7, and sample multiple trials (20 for GPT/Pa LM, 5 for Alpaca/Vicuna).