OpenICL

Why we included this project

OpenICL is aimed at researchers and graduate students who want to run in-context learning experiments without building every pipeline stage by hand. The framework turns the usual ICL workflow into a handful of Python objects: a DatasetReader loads and maps your data, a PromptTemplate controls how demonstrations and test inputs are formatted, a Retriever picks the in-context examples, and an Inferencer runs the model and scores the predictions. Several retrieval and inference strategies are included, so you can swap components and compare methods on the same task instead of writing glue code. LLaMA and self-consistency decoding are supported, and the paper plus the example notebooks give newcomers a clear way in. That makes it a practical base for systematic comparisons of language models and for fast prompting experiments.