Emergent inabilities? Inverse scaling over the course of pretraining
arXiv:2305.14681 · doi:10.18653/v1/2023.findings-emnlp.973
Abstract
Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task. We find 8 tasks on which Pythia 12B (Biderman et al., 2023) shows decreased performance over the course of training. Five of these tasks (TruthfulQA-MC1, TruthfulQA-MC2, Hindsight Neglect, Memo Trap, and Pattern Match Suppression) additionally show a consistent relationship whereby larger language models show a greater decrease in performance the more they are trained, despite showing standard (positive) scaling overall. This highlights the importance of testing performance at all relevant benchmarks any time models are trained on additional data, even if their overall performance improves
Accepted to Findings of EMNLP 2023
References in corpus (16)
- LLaMA: Open and Efficient Foundation Language Models
- Llama 2: Open Foundation and Fine-Tuned Chat Models
- PaLM: Scaling Language Modeling with Pathways
- Scaling Laws for Neural Language Models
- Emergent Abilities of Large Language Models
- LaMDA: Language Models for Dialog Applications
- Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models
- The Pile: An 800GB Dataset of Diverse Text for Language Modeling
- Galactica: A Large Language Model for Science
- Scaling Language Models: Methods, Analysis & Insights from Training Gopher
- GLaM: Efficient Scaling of Language Models with Mixture-of-Experts
- Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
- Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned
- Can Large Language Models Truly Understand Prompts? A Case Study with Negated Prompts
- Inverse Scaling: When Bigger Isn't Better
- Unified Scaling Laws for Routed Language Models