Analysis of 500 Hugging Face Model Cards Identifies Gaps in Model Heritage and Alignment Data for Open-Weight Models

The paper evaluates model cards against requirements for downstream governance of OWFMs. It documents incomplete reporting on training data lineage, fine-tuning steps, and post-training safety evaluations. Standard cards supply high-level capability claims but rarely include verifiable metrics on refusal rates or adversarial robustness after weight release.

Data from the 500-card sample reveals 72 percent omit alignment provenance details and 81 percent lack empirical behavior logs under misuse scenarios. These absences create an information gap that acceptable use policies and licenses must close. Standard open-source licenses such as Apache 2.0 provide no mechanism to enforce usage restrictions once weights are distributed.

The analysis connects the finding to prior transparency work on model reporting and to enforcement patterns observed after Llama 2 and Mistral weight releases. It shows that informational artifacts alone cannot substitute for normative constraints or legal instruments. Integrated artifacts must bind disclosure, usage rules, and license terms into a single enforceable package.

Repositories and standards bodies will need to prototype combined templates that embed AUP clauses directly into model card sections and reference specific license provisions. Pilot implementations on Hugging Face could test adoption rates and downstream compliance within 12 months.