Effective Structured Prompting by Meta-Learning and Representative Verbalizer

Authors: Weisen Jiang, Yu Zhang, James Kwok

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that Meta Prompter performs better than the recent state-of-the-arts and Rep Verb outperforms existing soft verbalizers.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, Southern University of Science and Technology 2Department of Computer Science and Engineering, Hong Kong University of Science and Technology 3Peng Cheng Laboratory.
Pseudocode Yes Algorithm 1 Representative Verbalizer (Rep Verb).Algorithm 2 Meta Prompter.
Open Source Code No The paper does not provide explicit statements or links indicating that the source code for the described methodology is open-source or publicly available.
Open Datasets Yes Following Chen et al. (2022), we perform few-shot classification on six popularly used data sets: (i) 20News (Lang, 1995), which contains informal discourses from news discussion forums of 20 topics; (ii) Amazon (He & Mc Auley, 2016), which consists of customer reviews from 24 products... (iii) Huff Post (Misra, 2022)... (iv) Reuters (Lewis, 1997)... (v) HWU64 (Liu et al., 2019)... (vi) Liu54 (Liu et al., 2019).
Dataset Splits Yes We use the meta-training/meta-validation/meta-testing splits provided in Chen et al. (2022).To prevent overfitting, we evaluate the meta-validation performance every 50 iteration and choose the checkpoint with the best meta-validation performance for meta-testing.
Hardware Specification Yes Experiments are run on a DGX station with 8 V100 32GB GPUs.
Software Dependencies No The paper mentions using 'bert-base-uncased' from Hugging Face and the 'Adam optimizer', but does not specify software versions for programming languages, libraries, or frameworks (e.g., Python version, PyTorch/TensorFlow version).
Experiment Setup Yes For Meta Prompter, hyperparameters K and Lp are chosen from {1, 2, 4, 8, 16, 32, 64} using the meta-validation set. For the base learner, α = 0.1, and J = 5 (resp. 15) at metatraining (resp. meta-validation or meta-testing). We train the prompt pool for T = 3, 000 iterations using the Adam optimizer (Kingma & Ba, 2015) with a learning rate of 0.001.