Create SIA_Dataset.py

SIA_Dataset.py

@@ -0,0 +1,193 @@
+"""
+SIABench — Custom Dataset Loading Script
+Place this file at the ROOT of the HuggingFace repo as SIA_Dataset.py
+"""
+
+import json
+import os
+import datasets
+
+_DESCRIPTION = """
+SIABench is a benchmark for evaluating LLMs on cybersecurity incident analysis tasks.
+It contains three sub-datasets:
+  - alert_triaging_tii: 100 alert triaging scenarios from TII-SSRC-23 (50 TP, 50 FP)
+  - alert_triaging_cic: Alert triaging scenarios from CIC datasets
+  - sia_dataset: 23 open-ended CTF network forensics scenarios
+"""
+
+_CITATION = """
+@inproceedings{siabench2025,
+  title={SIABENCH: A Benchmark for Evaluating LLMs on Cybersecurity Incident Analysis},
+  author={Jajodia, Sourov and others},
+  booktitle={PST 2025},
+  year={2025}
+}
+"""
+
+_BASE = "SIABench-main"
+
+
+class SIABenchConfig(datasets.BuilderConfig):
+    def __init__(self, data_dirs, **kwargs):
+        super().__init__(**kwargs)
+        self.data_dirs = data_dirs  # list of (split_name, folder_path) tuples
+
+
+class SIABench(datasets.GeneratorBasedBuilder):
+
+    BUILDER_CONFIGS = [
+
+        # ── Alert Triaging TII ──────────────────────────────────────────
+        SIABenchConfig(
+            name="alert_triaging_tii",
+            description="Alert triaging scenarios derived from TII-SSRC-23 dataset.",
+            data_dirs=[
+                ("tp", os.path.join(_BASE, "Alert_Triaging_Dataset_TII", "TP")),
+                ("fp", os.path.join(_BASE, "Alert_Triaging_Dataset_TII", "FP")),
+            ],
+        ),
+
+        # ── Alert Triaging CIC ──────────────────────────────────────────
+        SIABenchConfig(
+            name="alert_triaging_cic",
+            description="Alert triaging scenarios derived from CIC datasets.",
+            data_dirs=[
+                ("tp", os.path.join(_BASE, "Alert_Triaging_Dataset_CIC", "TP")),
+                ("fp", os.path.join(_BASE, "Alert_Triaging_Dataset_CIC", "FP")),
+            ],
+        ),
+
+        # ── SIA Dataset (CTF scenarios) ─────────────────────────────────
+        SIABenchConfig(
+            name="sia_dataset",
+            description="Open-ended CTF network forensics scenarios from CyberDefenders and BTLO.",
+            data_dirs=[
+                ("train", os.path.join(_BASE, "SIA_Dataset")),
+            ],
+        ),
+    ]
+
+    DEFAULT_CONFIG_NAME = "sia_dataset"
+
+    def _info(self):
+        if self.config.name in ("alert_triaging_tii", "alert_triaging_cic"):
+            features = datasets.Features({
+                "scenario_name": datasets.Value("string"),
+                "alert_name":    datasets.Value("string"),
+                "alert_type":    datasets.Value("string"),
+                "scenario":      datasets.Value("string"),
+                "alert":         datasets.Value("string"),
+                "tools_available": datasets.Sequence(datasets.Value("string")),
+                "files_available": datasets.Value("string"),
+                "instructions":  datasets.Value("string"),
+                "directory":     datasets.Value("string"),
+                "questions": datasets.Sequence({
+                    "question": datasets.Value("string"),
+                    "answer":   datasets.Value("string"),
+                }),
+            })
+        else:  # sia_dataset
+            features = datasets.Features({
+                "scenario_name":  datasets.Value("string"),
+                "source":         datasets.Value("string"),
+                "last_accessed":  datasets.Value("string"),
+                "writeup":        datasets.Value("string"),
+                "scenario":       datasets.Value("string"),
+                "task_category":  datasets.Value("string"),
+                "complexity":     datasets.Value("string"),
+                "tools_available": datasets.Sequence(datasets.Value("string")),
+                "files_available": datasets.Value("string"),
+                "instructions":   datasets.Value("string"),
+                "directory":      datasets.Value("string"),
+                "questions": datasets.Sequence({
+                    "question":          datasets.Value("string"),
+                    "answer":            datasets.Value("string"),
+                    "adversarial_tactic": datasets.Value("string"),
+                }),
+            })
+
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=features,
+            citation=_CITATION,
+            license="Apache-2.0",
+        )
+
+    def _split_generators(self, dl_manager):
+        return [
+            datasets.SplitGenerator(
+                name=split_name,
+                gen_kwargs={"folder": folder},
+            )
+            for split_name, folder in self.config.data_dirs
+        ]
+
+    def _generate_examples(self, folder):
+        if not os.path.exists(folder):
+            return
+
+        for fname in sorted(os.listdir(folder)):
+            if not fname.endswith(".json"):
+                continue
+
+            fpath = os.path.join(folder, fname)
+            with open(fpath, "r", encoding="utf-8") as f:
+                data = json.load(f)
+
+            scenarios = data.get("scenarios", [])
+
+            # ── Extract metadata block ──
+            meta = next(
+                (s.get("metadata") or {} for s in scenarios if "metadata" in s),
+                {}
+            )
+            # ── Extract sia_components block ──
+            comp = next(
+                (s.get("sia_components") or {} for s in scenarios if "sia_components" in s),
+                {}
+            )
+
+            if self.config.name in ("alert_triaging_tii", "alert_triaging_cic"):
+                questions = [
+                    {
+                        "question": q.get("question", ""),
+                        "answer":   q.get("answer", ""),
+                    }
+                    for q in comp.get("questions", [])
+                ]
+                yield fname, {
+                    "scenario_name":   meta.get("scenario_name", ""),
+                    "alert_name":      meta.get("alert_name", ""),
+                    "alert_type":      meta.get("alert_type", ""),
+                    "scenario":        comp.get("scenario", ""),
+                    "alert":           comp.get("alert", ""),
+                    "tools_available": comp.get("tools_available", []),
+                    "files_available": comp.get("files_available", ""),
+                    "instructions":    comp.get("instructions", ""),
+                    "directory":       comp.get("directory", ""),
+                    "questions":       questions,
+                }
+
+            else:  # sia_dataset
+                questions = [
+                    {
+                        "question":           q.get("question", ""),
+                        "answer":             q.get("answer", ""),
+                        "adversarial_tactic": q.get("adversarial_tactic", ""),
+                    }
+                    for q in comp.get("questions", [])
+                ]
+                yield fname, {
+                    "scenario_name":   meta.get("scenario_name", ""),
+                    "source":          meta.get("source", ""),
+                    "last_accessed":   meta.get("last_accessed", ""),
+                    "writeup":         meta.get("writeup", ""),
+                    "scenario":        comp.get("scenario", ""),
+                    "task_category":   comp.get("task_category", ""),
+                    "complexity":      comp.get("complexity", ""),
+                    "tools_available": comp.get("tools_available", []),
+                    "files_available": comp.get("files_available", ""),
+                    "instructions":    comp.get("instructions", ""),
+                    "directory":       comp.get("directory", ""),
+                    "questions":       questions,
+                }