Privacy-Preserving Instructions for Aligning Large Language Models

Authors: Da Yu, Peter Kairouz, Sewoong Oh, Zheng Xu

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

Reproducibility Variable Result LLM Response
Research Type Experimental In both supervised fine-tuning and reinforcement learning from human feedback, our extensive experiments demonstrate the high utility of the final set of synthetic instructions by showing comparable results to real instructions.
Researcher Affiliation Collaboration Da Yu Sun Yat-sen University, the work was done when Da Yu was an intern at Google Research. yuda3@mail2.sysu.edu.cn. Google Research, alphabetical order. {kairouz,sewoongo,xuzheng}@google.com.
Pseudocode Yes Algorithm 1 Train DP Generator to Synthesize Instructions. Algorithm 2 Resample Synthetic Data with DP Histogram.
Open Source Code Yes The code to reproduce the main findings in our paper is available at: https://github.com/google-research/google-research/tree/master/dp_instructions.
Open Datasets Yes We use the LMSYS-Chat-1M dataset as the private dataset (Zheng et al., 2023a)...For this, we take instructions from the FLAN dataset (Chung et al., 2022; Longpre et al., 2023).
Dataset Splits Yes The Chatbot Arena instructions are then divided into three subsets: 180,000 for the training set, 5,000 for the validation set, and the rest for the test set.
Hardware Specification Yes Our experiments use 32 Nvidia A100 40G GPUs.
Software Dependencies No The paper mentions several software components and libraries (e.g., LLaMA models, LoRA, DP-Adam, Adam optimizer, Sentence-T5-base, Faiss package, GPT-3.5-Turbo, GPT-4, phi-1.5, TRL library) but does not provide specific version numbers for them.
Experiment Setup Yes Our hyperparameter sweeps for DP fine-tuning and non-private fine-tuning are outlined in Table 9 and Table 10, respectively...The KL penalty coefficient is set as 0.2 and the number of optimisation steps per batch is 4. We use a batchsize of 128 and a learning rate of 1.41 × 10−5.