| --- |
| configs: |
| - config_name: alert_triaging_tii |
| data_files: |
| - split: tp |
| path: data/alert_triaging_tii_tp.jsonl |
| - split: fp |
| path: data/alert_triaging_tii_fp.jsonl |
|
|
| - config_name: alert_triaging_cic |
| data_files: |
| - split: test |
| path: data/alert_triaging_cic_test.jsonl |
|
|
| - config_name: sia_dataset |
| data_files: |
| - split: test |
| path: data/sia_dataset_test.jsonl |
|
|
| license: apache-2.0 |
| task_categories: |
| - question-answering |
| language: |
| - en |
| tags: |
| - cybersecurity |
| - incident-analysis |
| - LLM-benchmark |
| pretty_name: SIABench |
|
|
| --- |
| task_categories: |
| - question-answering |
| language: |
| - en |
| tags: |
| - cybersecurity |
| - incident-analysis |
| - LLM-benchmark |
| --- |
| |
| # SIABench — Security Incident Analysis Benchmark |
| |
| SIABench is a benchmark for evaluating LLMs on cybersecurity incident analysis tasks, developed at **Concordia University's Security Research Centre** in collaboration with **Defence Research and Development Canada (DRDC)**. |
| |
| ## Dataset Structure |
| |
| SIABENCH contains three sub-datasets accessible as separate configs: |
| |
| | Config | Description | Splits | |
| |---|---|---| |
| | `alert_triaging_tii` | 100 alert triaging scenarios (TII-SSRC-23 source) | `tp` (50), `fp` (50) | |
| | `alert_triaging_cic` | Alert triaging scenarios (CIC source) | `tp`, `fp` | |
| | `sia_dataset` | 25 open-ended CTF forensics scenarios | `train` | |
|
|
| ## Loading the Dataset |
|
|
| ```python |
| from datasets import load_dataset |
| |
| # Load alert triaging (TII) |
| tii = load_dataset("SIABench/SIA_Dataset", "alert_triaging_tii") |
| |
| # Load alert triaging (CIC) |
| cic = load_dataset("SIABench/SIA_Dataset", "alert_triaging_cic") |
| |
| # Load CTF investigation scenarios |
| sia = load_dataset("SIABench/SIA_Dataset", "sia_dataset") |
| ``` |
|
|
| ## Scenario Structure |
|
|
| Each JSON file contains two blocks: |
|
|
| **1. `metadata`** — scenario identification |
| - `scenario_name`: unique ID |
| - `alert_type` (TII/CIC only): `"True"` (TP) or `"False"` (FP) |
|
|
| **2. `sia_components`** — the LLM task |
| - `scenario`: natural language description of the incident |
| - `tools_available`: CLI tools the LLM agent can invoke |
| - `files_available`: linked PCAP file (`capture.pcap`) |
| - `directory`: set to `"(directory_of_the_file)"` |
| - `questions`: Q&A pairs (with `adversarial_tactic` in SIA_Dataset) |
