Featuring Every Eval Ever Results on Hugging Face Model Pages

Source: HuggingFace Date: 2026-06-30 URL: https://huggingface.co/blog/eee-community-evals

Summary

Hugging Face integrates with Every Eval Ever (EEE), a standardized JSON schema for benchmark results, so model cards can surface scores that link back to full reproducibility records (harness version, generation config, access method). The stated problem is real fragmentation — the post cites LLaMA 65B reporting both 63.7 and 48.8 on MMLU depending on who ran it — and the combined datastore already holds ~229,000 results across 22,000+ models and 2,200 benchmarks.

Implications

Feeds the open-weight clock most directly: this is exactly the kind of infrastructure needed to make the “trending != shipping” verification discipline (checking createdAt, provenance, and actual benchmark methodology rather than a headline number) cheaper to do at scale. It’s also a small dev tooling signal — if the converter and schema get adopted broadly, “read the eval methodology” becomes a one-click model-card action instead of a paper-hunting exercise, which lowers the cost of the skepticism this project already applies by hand.