🧑‍🔬 Alexander Ratner

Alexander Ratner's papers here study how large language models can supply the training signal for other models. Both works, from 2023, ask how to get useful supervision out of an LLM while lowering the cost of labelled data or of serving a large model. One uses machine-written explanations to train compact models. The other examines how the wording of generation prompts shapes the variety and skew of synthetic classification data. The work sits within Language models & pretraining.

The most cited paper, Distilling Step-by-Step! (2023), prompts PaLM 540B with few-shot chain-of-thought examples. For every example in the training set, the large model returns a label together with a written justification. A smaller T5 model is then taught two jobs, marked by separate input prefixes: predicting the label and producing the explanation. Only label prediction runs at deployment, so the extra training objective adds no serving cost. The second paper, Large Language Model as Attributed Training Data Generator (2023), shows that simple class-conditional prompts produce repetitive and biased datasets. It proposes AttrPrompt instead. With some human guidance, the LLM helps define attributes such as subtopic, style, length and location. Random mixes of their values then fill the generation prompts. The method is evaluated with ChatGPT on four topic-classification benchmarks that have many, unevenly sized classes.

Both of these papers are also the newest here. Together they explore LLM-generated rationales and synthetic examples as a way to reduce reliance on human annotation. Ratner is currently at the University of Washington in Seattle, after earlier work at Stanford University from 2014 to 2020. Ranjay Krishna is a co-author on two papers. Other collaborators include Yu Meng, Yue Yu and Jieyu Zhang.

Alexander Ratner has 2 papers in our library, in 2023, most of them on Language models & pretraining; often with Ranjay Krishna, Chao Zhang, Jiaming Shen.