Continual-Mega/ADCT

🧠 Continual-MEGA: A Large-scale Benchmark for Generalizable Continual Anomaly Detection

This repository contains the evaluation code for the Continual-Mega benchmark, submitted to ICLR 2026.

We provide the checkpoint files of the proposed model along with the evaluation code for the Continual-Mega benchmark.

The training code will be made publicly available at a later date.

🔗 Codebase: Continual-Mega/Continual-Mega

---

🚀 Overview

Continual-MEGA introduces a realistic and large-scale benchmark for continual anomaly detection that emphasizes generalizability across domains and tasks.

The benchmark features:

• ✅ Diverse anomaly types across domains
• 🔁 Class-incremental continual learning setup
• 📈 A large-scale evaluation protocol surpassing previous benchmarks

This repository hosts pretrained model checkpoints used in various scenarios defined in the benchmark.

---

📦 Available Checkpoints

Model Name	Scenario	Task	Description
scenario2/prompt_maker	Scenario 2	Base	Prompt maker trained on Scenario 2 base classes
scenario2/adapters_base	Scenario 2	Base	Adapter trained on Scenario 2 base classes
scenario2/30classes/adapters_task1	Scenario 2	Task 1 (30 classes)	Adapter trained on Task 1 (30 classes) in Scenario 2
scenario2/30classes/adapters_task2	Scenario 2	Task 2 (30 classes)	Adapter trained on Task 2 (30 classes) in Scenario 2
(More to come)	–	–	–

---

🛠 Usage Example

Continual Setting Evaluation

sh eval_continual.sh

Zero-Shot Evaluation

sh eval_zero.s