Hugging Face and the EvalEval Coalition have made Every Eval Ever (EEE) evaluation records intercompatible with Hugging Face Community Evals, displaying standardized benchmark scores directly on model pages.
Hugging Face reported that evaluation scores often vary significantly across sources due to unreported settings, pointing to LLaMA 65B scoring both 63.7 and 48.8 on MMLU in separate reports. Both EEE and Community Evals originally launched in February 2026 to resolve these reporting gaps across first-party and third-party AI evaluators.
The EEE datastore on Hugging Face now holds around 229,000 evaluation results across more than 22,000 models and 2,200 benchmarks, converted from 31 different reporting formats. Hugging Face noted that reproducing those benchmark runs from scratch would cost hundreds of thousands of dollars. EEE structures each result into a single JSON schema that records metric definitions alongside optional companion JSONL files containing per-sample output logs.
Integration and verified checkmarks
A converter script automates cross-posting by translating EEE records into YAML files stored under .eval_results inside Hugging Face model repositories. Submissions made from an organization's official Hugging Face account display a verified checkmark on EvalEval. Corresponding scores on the Hugging Face Hub display badges linking back to the full EEE record, including generation configurations and reproducibility notes.
The converter tool currently supports four official benchmarks: MMLU-Pro, GPQA, Humanity's Last Exam, and GSM8K. During execution, the tool audits target model repositories for existing files, marking results as already_present, score_conflict, missing_hf_model, or ready. It writes local previews and requires users to type OPEN PRS before opening pull requests on the Hub.
