Industrial anomaly understanding

The MMAD Leaderboard

Explore MMAD results and their evaluation settings.
Every result comes with a source. All reported protocols together, with filters for their differences.

Last literature review

Loading…Read the evidence ledger ↗
Standard protocol:1-shotNo MMAD adaptationNo added knowledge / RAGFull eligibility rules ↓

Other industrial training is allowed. The main track includes industrial specialists trained on disjoint datasets and general-purpose backbones. MMAD / constituent-source adaptation is excluded from the standard protocol. See training sources and model decisions ↓

—
Reported resultsAll settings · filter below
7
Inspection tasksFull MMAD evaluation
4
Source datasetsAll domains required

Results & evidence

Loading source…
Loading results…
Paper-reported results, with primary-source protocol checks. Loading the latest ledger…

Loading leaderboard…

Standard rank applies only to standard 1-shot rows · reported scores use the labeled settingsDataset revisions & provenance ↗

A comparable main track

One normal reference, the full benchmark, and no task-specific access to MMAD beyond the evaluation input.

Read all five eligibility rules
  1. Loading the reviewed protocol…

“Training-free” does not rule out extra normal memory banks or retrieved domain text. Unclear provenance remains pending. A separate track describes a different protocol, not an allegation of misconduct.

Help keep the benchmark current

Submit a paper, a result, or a correction with the checkpoint, dataset revision, exact prompt, all component training sources, and reference budget. We review the evidence before adding a ranked result.

Submit a result ↗

Submission requirements ↗