229,000 benchmark scores now link directly to Hugging Face model cards

The mess of AI evaluation just got a little easier to navigate. The EvalEval Coalition’s EEE project, launched in February 2026, has now built a direct pipeline to Hugging Face’s Community Evals, meaning those 229,000 structured benchmark results won’t just sit in a separate datastore — they’ll show up right on model pages. If you’ve ever wondered why LLaMA 65B scores anywhere from 48.8 to 63.7 on MMLU depending on who you ask, that’s the exact problem this integration is trying to solve.

The new converter takes a full EEE JSON record — which captures who ran the test, how they accessed the model, what generation settings they used, and what the metric actually means — and spits out the small YAML files Hugging Face needs for its Community Evals system. Nobody has to maintain two formats by hand anymore. A researcher runs an eval, dumps the data into EEE, and with one extra step it lands on the model card with a verified badge linking back to the full record, including per-sample outputs and reproducibility notes.

The scale here is worth pausing on. The EEE datastore pulls from 31 different reporting formats — papers, leaderboards, harness logs, blog posts — and covers over 22,000 models and 2,200 benchmarks. Reproducing those runs from scratch would cost somewhere in the hundreds of thousands of dollars. So the coalition’s argument isn’t just about convenience; it’s about not burning money to regenerate data someone already paid for.

For model authors submitting through their official Hugging Face accounts, there’s an added perk: their results get a verified checkmark on EvalEval, signaling that the numbers come straight from the source rather than some anonymous third party. The integration doesn’t solve the deeper problem of evaluation design — bad benchmarks are still bad benchmarks — but it chips away at the trust gap between what a model claims and what you can actually verify. That’s more than most evaluation reporting standards have managed so far.