Paper | Project page | Code
Agent Retrieval Bench (ARB) evaluates the context-acquisition layer of coding
agents: given a real workflow signal and a repository at a frozen base commit,
can a retriever find the files an agent needs to read next, or correctly decide
that the repository contains no useful local context?
ARB differs from generic semantic code search in its relevance definition.
Relevant files are selected by the next need of the coding… See the full description on the dataset page:
https://huggingface.co/datasets/eyuansu71/agent_retrieval_bench.
