Download build_items.py from MK4-Research/GuardBench: direct link, hf CLI and curl.
- Browser
- Download file 3.04 kB
-
https://huggingface.co/datasets/MK4-Research/GuardBench/resolve/main/build_items.py
- Command line
-
hf download hf://datasets/MK4-Research/GuardBench/build_items.py
-
curl -L -o build_items.py https://huggingface.co/datasets/MK4-Research/GuardBench/resolve/main/build_items.py
3.04 kB
| #!/usr/bin/env python3 | |
| """Build GuardBench items: 20 patterns x 6 shapes. | |
| Ground truth is set by the delta, not by judgement. The guard is always at distance from the | |
| sink - a helper above, a decorator, a module constant, or a branch several lines up - because a | |
| guard adjacent to its sink cannot be missed and that is exactly why FBE could not measure this. | |
| """ | |
| import json, ast, os, sys, hashlib | |
| from collections import Counter | |
| HERE = os.path.dirname(os.path.abspath(__file__)) | |
| sys.path.insert(0, HERE) | |
| import build_items_v1 as V1 | |
| from patterns_v2 import P2 | |
| from covers_alt_v1 import COVERS_ALT | |
| SHAPES = ["none", "covers", "covers_alt", "wrong_value", "irrelevant", "elsewhere"] | |
| LABEL = {"none":"unsafe", "covers":"safe", "covers_alt":"safe", | |
| "wrong_value":"unsafe", "irrelevant":"unsafe", "elsewhere":"unsafe"} | |
| WHY = { | |
| "none": "no guard is applied on this path", | |
| "covers": "the guard is applied to the tainted value and addresses the weakness", | |
| "covers_alt": "a second correct implementation, defended a different way", | |
| "wrong_value": "the guard is applied to a sibling value, not the one that reaches the sink", | |
| "irrelevant": "a guard is applied to the right value but does not address this weakness", | |
| "elsewhere": "the guard exists and is used by a neighbouring function, not on this path", | |
| } | |
| def build(): | |
| out, seen = [], set() | |
| def add(name, cls, shape, code): | |
| h = hashlib.sha1(code.encode()).hexdigest()[:12] | |
| if h in seen: return | |
| seen.add(h) | |
| out.append({"id": f"{name}_{shape}", "pattern": name, "shape": shape, | |
| "label": LABEL[shape], "vuln_class": cls if LABEL[shape]=="unsafe" else None, | |
| "why": WHY[shape], "code": code}) | |
| for fn in V1.P: # 8 v1 patterns, 5 shapes + covers_alt | |
| cls, *variants = fn(None) | |
| for shape, code in zip(["none","covers","wrong_value","irrelevant","elsewhere"], variants): | |
| add(fn.__name__, cls, shape, code) | |
| alt = COVERS_ALT.get(fn.__name__) | |
| if alt: add(fn.__name__, cls, "covers_alt", alt) | |
| for fn in P2: # 12 v2 patterns, 6 shapes | |
| cls, d = fn(None) | |
| for shape in SHAPES: | |
| if shape in d: add(fn.__name__, cls, shape, d[shape]) | |
| return out | |
| if __name__ == "__main__": | |
| rows = build() | |
| bad = [] | |
| for r in rows: | |
| try: ast.parse(r["code"]) | |
| except SyntaxError as e: bad.append((r["id"], str(e))) | |
| p = os.path.join(HERE, "items.jsonl") | |
| with open(p, "w") as fh: | |
| for r in rows: fh.write(json.dumps(r)+"\n") | |
| lab = Counter(r["label"] for r in rows) | |
| print(f"items: {len(rows)} -> {p}") | |
| print("patterns:", len({r['pattern'] for r in rows})) | |
| print("shape :", dict(Counter(r['shape'] for r in rows))) | |
| print("label :", dict(lab), f" safe share {lab['safe']/len(rows):.0%}") | |
| print(f"classes: {len({r['vuln_class'] for r in rows if r['vuln_class']})}") | |
| print(f"syntax errors: {len(bad)}") | |
| for i,e in bad[:6]: print(" ", i, e) | |