AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

The Stability-Efficiency Dilemma: Investigating Sequence Length Warmup for Training GPT Models

Conglong Li, Minjia Zhang, Yuxiong He

NeurIPS 2022 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 18 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add sequence length warmup (SLW) at batch size 4KGPT-2 1.5B · WikiText-103 zero-shotPPL: 14.76 -> 13.72Improved
add SLW at batch size 4KGPT-2 117M · WikiText-103 zero-shotPPL: 28.09 -> 27.15Improved
add SLW at batch size 512GPT-2 1.5B · WikiText-103 zero-shotPPL: 13.89 -> 13.88No clear effect
add SLW with sequence length 2KGPT-2 117M · WikiText-103 zero-shotPPL: 28.19 -> 26.03Improved
replace Shortformer two-stage short-then-full sequences with SLWGPT-2 1.5B bsz4K · WikiText-103 zero-shotPPL: 14.14 -> 13.72Improved
replace batch size warmup with SLWGPT-2 1.5B bsz4K · WikiText-103 zero-shotPPL: 14.21 -> 13.72Improved
add SLW enabling 40x LR, 30B tokensGPT-3 125M · 11-task zero-shot averageaverage accuracy: 29.8 -> 31.1Improved
add SLW to GPT-3 1.3B at 300B tokensGPT-3 1.3B · zero-shot average accuracyaverage accuracy: 41.6 -> 41.9Improved
use discrete 2-stage pacing function instead of linear step-wise (ablation)SLW · GPT-2 pretrainingstability/convergenceHurt
use root or adaptive pacing function instead of linear (ablation)SLW · GPT-2 pretrainingconvergenceNo clear effect
vary SLW duration T within reasonable range (ablation)SLW on GPT-2 117M · WikiText/LAMBADAaccuracyNo clear effect
increase gradient clipping strength instead of SLW (ablation)GPT-2 baseline · GPT-2 pretrainingtraining stabilityNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy