Structured Voronoi Sampling

Authors: Afra Amini, Li Du, Ryan Cotterell

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In an experimental setup where the reference distribution is known, we show that the empirical distribution of SVS samples is closer to the reference distribution compared to alternative sampling schemes.
Researcher Affiliation Academia Afra Amini1 Li Du2 Ryan Cotterell1 1ETH Zürich 2Johns Hopkins University {afra.amini, ryan.cotterell}@inf.ethz.ch leodu@cs.jhu.edu
Pseudocode Yes Algorithm 1 HMC, Algorithm 2 Langevin Dynamics, Algorithm 3 MUCOLA, Algorithm 4 Structured Voronoi Sampling, Algorithm 5 REFRACTREFLECT, Algorithm 6 Find Discontinuity
Open Source Code Yes https://github.com/Afra Amini/svs
Open Datasets Yes The underlying LM is a finetuned GPT-210 on E2E dataset [34]; see App. G for dataset statistics. This dataset is made available under the CC BY-SA 4.0 license.
Dataset Splits Yes Table 2: Number of restaurant reviews in each split and food type. train 2929... valid 1489... test 492...
Hardware Specification Yes All experiments are done on a single A100-40GB GPU. All classifiers are trained and tested on a single gtx_1080_ti GPU with approximately 2 hours of total computational budget.
Software Dependencies No The paper mentions using a 'gpt2 checkpoint from the Huggingface library [47]' but does not specify version numbers for the Huggingface library itself or other core software dependencies like Python or PyTorch, which would be necessary for full reproducibility.
Experiment Setup Yes Hyperparameters for each experiment are reported in Table 4. Following prior work [46], in algorithms based on Langevin dynamics, we apply an exponential decay to the step size by decreasing it to 0.05 after 500 steps. In all settings, we take 500 burn-in steps.