majority label bias
The tendency of a few-shot prompt to favour the label that appears most often, or last, among its examples.
In more detail
Zhao and colleagues described it in 2021 in "Calibrate Before Use", along with a bias towards common tokens. Balance the labels in your examples, vary their order and measure on held-out data.
Related terms
- few-shot prompting — Including a few worked input and output examples in the prompt to show the model the task and format.
Taught in
- Zero-Shot And Few-Shot — AI Engineering: LLM Apps, RAG & Agents
References
- Zhao et al. (2021): Calibrate Before Use: Improving Few-Shot Performance of Language Models
- Lu et al. (2022): Fantastically Ordered Prompts and Where to Find Them