The metadata from Papers-with-code, OpenML and Huggingface follow different data structures and nomenclatures. For example, the concept "Model" is referred to as "Methods" in Papers-with-Code, "Flow" in OpenML, and "Model" in Hugging Face. We propose the Common Metadata Ontology (CMO), a unifying schema to integrate diverse data structures from Papers-with-Code, OpenML, and Hugging Face. Built on the [Common Metadata Framework (CMF)](https://github.com/HewlettPackard/cmf), CMO ensures interoperability of metadata, enabling knowledge discovery, search, and reasoning capabilities. The overview of CMO is shown in Figure 1 below. The novel features of CMO are as follows: <br>
* (i) following a pipeline-centric approach, similar to CMF, to capture multiple experimentation runs for each stage (train, test, validation) with parameter settings, facilitating the identification of the best execution path
* (ii) modularity that allows distributed experiments and parallel logging of pipeline metadata, enabling seamless metadata capture across different teams and machines
* (iii) support for additional semantic and statistical properties that can be extracted, computed, or generated from entity names (e.g., identifying tasks as image-based, text-based, or audio-based); and
* (iv) support for multimodal properties, including text and vector embeddings of entity names, to enable keyword and approximate search.
<br>
A detailed overview of CMO and its properties can be found below. MLSchema and MLFlow adopt a model-centric approach. When building a pipeline, say, entity extraction from semi-structured electronic health records and testing it with multiple models, MLSchema and MLFlow require creating several instances — one for each model — to record metadata. In contrast, CMO allows all models, variations, hyperparameters, metrics, and datasets to be documented as a single instance, facilitating a scaleable and flexible metadata recording process by taking a holistic view of the entire pipeline. Hence, CMO builds upon the principles of CMF. For example,