What it is
An auditable research project on risk detection for Chinese social-media posts that combine images and text.
What it does: Runs a 4-bit Qwen3-VL-32B with a language-side LoRA locally to produce a strict seven-field JSON of labels and evidence, strips labels and identifiers, and asks Jev for a fixed-policy decision that maps to risk bands for human review. TF-IDF baselines run alongside and can be combined with an OR rule.
Why it belongs in the showcase: It places Jev inside a larger moderation pipeline and publishes the negative result: on the frozen test set, the Jev plan did not beat a strong text baseline.
Caveats: Small test set, one label, no public data or weights.
Evidence
Public MIT README (Chinese, with an English version); frozen 170-item test set (47 positive for the dataset's misogyny label); best test Macro-F1 90.36% from TF-IDF + logistic regression; the pre-registered "text OR multimodal + Jev" plan scored 87.25%, below the text baseline; on the validation set, text-only Jev 69.42% vs LoRA evidence + Jev 84.25% Macro-F1 at t=0.40; only the transcript and six evidence fields are sent to Jev.
Caveats
- Small test set; the positive class is one dataset label (misogyny), not all content risk
- Metrics are summaries from the author's experiment logs; the public repo does not rerun GPU or cloud evaluations, and no dataset, predictions or weights are published
- Not an automatic takedown system; Jev results here depend on Qwen evidence and are not a pure Jev baseline
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.