This document provides an overview of the supplementary materials submitted in support of our paper:

Paper Title: Delving into Large Language Models for Effective Time-Series Anomaly Detection

• This repository is built upon AnomLLM, and therefore must be set up using the AnomLLM environment.
  Please follow the installation instructions provided in the AnomLLM repository before proceeding.

• Dataset Download

  • AnomLLM: Downloaad "anomllm.zip" in the provied link in README of https://github.com/rose-stl-lab/anomllm
  • TSB-AD-U: Download "Datasets" directory in https://github.com/TheDatumOrg/TSB-AD/tree/main/Datasets

• For Qwen and InternVL models, we use LMDeploy.

• File Structure

LLM-TSAD/
│
├── ...
├── credentials.yml               # For online API
├── data
     ├── synthetic                # For AnomLLM benchmark
├── TSB-AD
     ├── Datasets                 # For TSB-AD-U benchmark
└── README.md                     # This file

1. Run online api (For convenience, we have saved the results of the previous run in the ./results/ directory. Therefore, you can proceed directly to step 2.)

python src/LLM-TSAD-AnomLLM_api.py --model gemini-1.5-flash --data trend --variant 0shot-text-vision

2. Aggregate evaluation results

python ./src/result_agg_by_model.py --model gemini-1.5-flash --benchmark anomllm

1. Run online api (For convenience, we have saved the results of the previous run in the ./results/ directory. Therefore, you can proceed directly to step 2.)

python src/LLM-TSAD-TSB_api.py --model gemini-1.5-flash --datadir ./TSB-AD/Datasets

2. Aggregate evaluation results

python ./src/result_agg_by_model.py --model gemini-1.5-flash  --benchmark tsb-ad-u