Mesa-Extrapolation: A Weave Position Encoding Method for Enhanced Extrapolation in LLMs

Authors: Xin Ma, Yang Liu, Jingjing Liu, Xiaoxu Ma

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments validate the effectiveness of Mesa-Extrapolation, demonstrating its potential as a scalable solution to enhancing LLMs applicative reach.
Researcher Affiliation Collaboration Xin Ma1, Yang Liu2,3 , Jingjing Liu2, Xiaoxu Ma1 1Digital Research Institute, Enn Group, Beijing, China 2Institute for AI Industry Research, Tsinghua University, Beijing, China 3Shanghai Artificial Intelligence Laboratory, China
Pseudocode Yes Algorithm 1 Mesa-Extrapolation Algorithm
Open Source Code Yes Our code is available at https: //github.com/soacker/Mesa-Extrapolation.
Open Datasets Yes We choose Gov Report Huang et al. (2021), Pile Gao et al. (2020), Long Bench Bai et al. (2023), and Long Eval Krishna et al. (2023) datasets, and also generate a passkey dataset, which has been integrated in the code warehouse.
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning.
Hardware Specification Yes We use a 2x A800 80GB NVIDIA GPU server as the experimental environment and adopt the Py Torch framework.
Software Dependencies No The paper only mentions "Py Torch framework" without providing specific version numbers for software dependencies.
Experiment Setup Yes In general, we set F = 100, Mmax = 200 and L = 512. Additionally, Stair PE is primarily employed in the manipulation of the last chunk... In Equ.3, we generally set the extrapolated position N = 512 and set the extrapolated width E = 50.