Cloud service composition involves combining multiple autonomous services to deliver new value-added services with enhanced functionalities. Existing solutions rely mostly on structured or semi-structured requests, requiring technical expertise and limiting accessibility for non-expert users. With the advent of Large Language Models (LLMs), it is now feasible to interpret end-user intents expressed in Natural Language (NL) and automatically generate corresponding service compositions. This marks a shift toward automating cloud service composition based on unstructured requests. To enable functional composition from such requests, their interpretation is a crucial step. It enables the identification of both explicit and implicit needs, mapping them to relevant services and automatically constructing a logical, provider-agnostic composition (i.e., relevant services with their dependencies). The latter serves as the foundation for downstream tasks—including service discovery, selection, and execution code generation—ultimately producing a deployable service composition. However, this critical interpretation task is often overlooked in existing research, primarily due to the lack of appropriate datasets tailored for cloud service composition from natural language requests. In this paper, we first propose an LLM-assisted method for constructing a benchmark dataset that captures diverse user profiles and varying levels of NL requests completeness. Each request is paired with a corresponding provider-independent composition, formally represented using the standard TOSCA specification language. Second, using this dataset, we evaluate the performance of both open-source and proprietary LLMs on the interpretation of technically diverse and completeness-varying requests.