PaLM: Scaling Language Modeling with Pathways
Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning, which drastically reduces the number of task-specific training examples needed to adapt the model to a particular application. To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model PaLM. We trained PaLM on 6144 TPU v4 chips using Pathways, a new ML system which enables highly efficient training across multiple TPU Pods. We demonstrate continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. On a number of these tasks, PaLM 540B achieves breakthrough performance, outperforming the finetuned state-of-the-art on a suite of multi-step reasoning tasks, and outperforming average human performance on the recently released BIG-bench benchmark. A significant number of BIG-bench tasks showed discontinuous improvements from model scale, meaning that performance steeply increased as we scaled to our largest model. PaLM also has strong capabilities in multilingual tasks and source code generation, which we demonstrate on a wide array of benchmarks. We additionally provide a comprehensive analysis on bias and toxicity, and study the extent of training data memorization with respect to model scale. Finally, we discuss the ethical considerations related to large language models and discuss potential mitigation strategies.
PDF Abstract Google Research 2022 PDFCode
Tasks
Auto Debugging
Code Generation
Common Sense Reasoning
Coreference Resolution
Cross-Lingual Question Answering
Few-Shot Learning
Hindu Knowledge
Known Unknowns
Language Modelling
Logic Grid Puzzle
Memorization
Multi-task Language Understanding
Novel Concepts
Question Answering
Reading Comprehension
StrategyQA
Winowhy
GLUE
Natural Questions
SuperGLUE
TriviaQA
RACE
WSC
COPA
BoolQ
WebQuestions
WinoGrande
HellaSwag
TyDi QA
MultiRC
LAMBADA
ReCoRD
HumanEval
BIG-bench
MMLU
GEM
TyDiQA-GoldP
| Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Benchmark |
|---|---|---|---|---|---|---|
| Novel Concepts | BIG-bench | PaLM-540B (few-shot, k=5) | Accuracy | 71.9 | # 1 | |
| StrategyQA | BIG-bench | PaLM-62B (few-shot, k=5) | Accuracy | 65.4 | # 3 | |
| Winowhy | BIG-bench | PaLM-540B (few-shot, k=5) | Accuracy | 65.9 | # 1 | |
| Winowhy | BIG-bench | PaLM-62B (few-shot, k=5) | Accuracy | 61.0 | # 3 | |
| Known Unknowns | BIG-bench | PaLM-540B (few-shot, k=5) | Accuracy | 73.9 | # 1 | |
| Hindu Knowledge | BIG-bench | PaLM-540B (few-shot, k=5) | Accuracy | 95.4 | # 1 | |
| Hindu Knowledge | BIG-bench | PaLM-62B (few-shot, k=5) | Accuracy | 77.7 | # 3 | |
| Logic Grid Puzzle | BIG-bench | PaLM-540B (few-shot, k=5) | Accuracy | 42.4 | # 2 | |
| Logic Grid Puzzle | BIG-bench | PaLM-62B (few-shot, k=5) | Accuracy | 36.5 | # 3 | |
| Novel Concepts | BIG-bench | PaLM-62B (few-shot, k=5) | Accuracy | 59.4 | # 3 | |
| StrategyQA | BIG-bench | PaLM-540B (few-shot, k=5) | Accuracy | 73.9 | # 1 | |
| Auto Debugging | Big-bench Lite | PaLM 8B (few-shot, k=5) | Exact string match | 14.7 | # 3 | |
| Auto Debugging | Big-bench Lite | PaLM 540B (few-shot, k=5) | Exact string match | 38.2 | # 1 | |
| Auto Debugging | Big-bench Lite | PaLM 62B (few-shot, k=5) | Exact string match | 38.2 | # 1 | |
| Question Answering | BoolQ | PaLM 540B (finetuned) | Accuracy | 92.2 | # 2 | |
| Natural Language Inference | CommitmentBank | PaLM 540B (finetuned) | F1 | 100 | # 1 | |
| Accuracy | 100 | # 1 | ||||
| Question Answering | COPA | PaLM 540B (finetuned) | Accuracy | 100 | # 1 | |
| Extreme Summarization | GEM-XSum | PaLM (finetuning)-540B | ROUGE-2 | 21.2 | # 2 | |
| Parameters | 540 B | # 2 | ||||
| Extreme Summarization | GEM-XSum | T5-XXL | ROUGE-2 | 21.0 | # 3 | |
| Extreme Summarization | GEM-XSum | PaLM (finetuning)-62B | ROUGE-2 | 18.5 | # 4 | |
| Parameters | 62 B | # 3 | ||||
| Sentence Completion | HellaSwag | PaLM-540B (Few-Shot) | Accuracy | 83.8 | # 5 | |
| Sentence Completion | HellaSwag | PaLM-540B (Zero-Shot) | Accuracy | 83.4 | # 7 | |
| Sentence Completion | HellaSwag | PaLM-540B (One-Shot) | Accuracy | 83.6 | # 6 | |
| Code Generation | HumanEval | PaLM 8B | Pass@1 | 3.6 | # 20 | |
| Pass@100 | 18.7 | # 13 | ||||
| Code Generation | HumanEval | PaLM-cont 62B | Pass@1 | 23.7 | # 9 | |
| Code Generation | HumanEval | PaLM 62B | Pass@1 | 15.9 | # 15 | |
| Pass@100 | 46.3 | # 9 | ||||
| Code Generation | HumanEval | PaLM 540B | Pass@1 | 26.2 | # 8 | |
| Pass@100 | 76.2 | # 2 | ||||
| Language Modelling | LAMBADA | PaLM-540B (Few-Shot) | Accuracy | 89.7 | # 1 | |
| Language Modelling | LAMBADA | PaLM-540B (One-Shot) | Accuracy | 81.8 | # 5 | |
| Language Modelling | LAMBADA | PaLM-540B (Zero-Shot) | Accuracy | 77.9 | # 10 | |
| Multi-task Language Understanding | MGSM | PaLM 540B | Average (%) | 55.0 | # 4 | |
| Multi-task Language Understanding | MMLU | PaLM 540B (few-shot, k=5) | Humanities | 77.0 | # 1 | |
| Average (%) | 69.3 | # 12 | ||||
| Parameters (Billions) | 540 | # 40 | ||||
| STEM | 55.6 | # 7 | ||||
| Social Sciences | 81.0 | # 1 | ||||
| Other | 69.6 | # 5 | ||||
| Tokens (Billions) | 780 | # 5 | ||||
| Question Answering | MultiRC | PaLM 540B (finetuned) | F1 | 90.1 | # 1 | |
| EM | 69.2 | # 1 | ||||
| Question Answering | Natural Questions | PaLM-540B (Zero-Shot) | EM | 21.2 | # 29 | |
| Question Answering | Natural Questions | PaLM-540B (One-Shot) | EM | 29.3 | # 24 | |
| Question Answering | Natural Questions | PaLM-540B (Few-Shot, k=64) | EM | 39.6 | # 18 | |
| Question Answering | OBQA | PaLM 540B (zero-shot) | Accuracy | 53.4 | # 7 | |
| Question Answering | OBQA | PaLM 62B (zero-shot) | Accuracy | 50.4 | # 8 | |
| Reading Comprehension | RACE | PaLM 62B (zero-shot) | Accuracy (High) | 47.5 | # 10 | |
| Accuracy (Middle) | 64.3 | # 9 | ||||
| Reading Comprehension | RACE | PaLM 540B (zero-shot) | Accuracy (High) | 49.1 | # 8 | |
| Accuracy (Middle) | 68.1 | # 7 | ||||
| Reading Comprehension | RACE | PaLM 8B (zero-shot) | Accuracy (High) | 42.3 | # 14 | |
| Accuracy (Middle) | 57.9 | # 14 | ||||
| Common Sense Reasoning | ReCoRD | PaLM 540B (finetuned) | F1 | 94.6 | # 1 | |
| EM | 94.0 | # 2 | ||||
| Natural Language Inference | RTE | PaLM 540B (finetuned) | Accuracy | 95.7% | # 1 | |
| Question Answering | TriviaQA | PaLM-540B (Few-Shot) | EM | 81.4 | # 1 | |
| Question Answering | TriviaQA | PaLM-540B (Zero-Shot) | EM | 76.9 | # 4 | |
| Question Answering | TriviaQA | PaLM-540B (One-Shot) | EM | 81.4 | # 1 | |
| Cross-Lingual Question Answering | TyDiQA-GoldP | PaLM-540B (CoT) | EM | 52.9 | # 7 | |
| Question Answering | WebQuestions | PaLM-540B (Few-Shot) | EM | 43.5 | # 5 | |
| Question Answering | WebQuestions | PaLM-540B (Zero-Shot) | EM | 10.6 | # 14 | |
| Question Answering | WebQuestions | PaLM-540B (One-Shot) | EM | 22.6 | # 11 | |
| Common Sense Reasoning | WinoGrande | PaLM 62B (zero-shot) | Accuracy | 77.0 | # 4 | |
| Common Sense Reasoning | WinoGrande | PaLM-cont 62B (zero-shot) | Accuracy | 77.0 | # 4 | |
| Common Sense Reasoning | WinoGrande | PaLM 540B (zero-shot) | Accuracy | 81.1 | # 3 | |
| Word Sense Disambiguation | Words in Context | PaLM 540B (finetuned) | Accuracy | 78.8 | # 2 | |
| Coreference Resolution | WSC | PaLM 540B | Accuracy | 89.5 | # 1 |
