use XDG paths for configuration data and caching

Support using [XDG ver 0.8](https://specifications.freedesktop.org/basedir-spec/basedir-spec-latest.html) for project data.

Specifically support:
| ENV VAR          | DEFAULT            |
|------------------|--------------------|
| $XDG_DATA_HOME   | $HOME/.local/share |
| $XDG_CONFIG_HOME | $HOME/.config      |
| $XDG_CACHE_HOME  | $HOME/.cache       |

Project name `garak` is appended to each location.

This is represents the followina breaking changes to project expecations:
* report_prefix passed either at the command line or as config file option
  * set filename values only
  * no longer overrides report_dir
* report_dir passed as a config file option
  * when provided as a relative path will be prepend with `<xdg_data_home>/garak`
  * provided as an absolute path will be used as the output directory
* default `user/site` configuration file `garak.site.yaml` has moved
  * previously `<basedir>/garak.site.yaml`
  * updated location `<xdg_config_home>/garak/garak.site.yaml`

Additional changes (not considered breaking changes):
* nltk data is placed in <xdg_cache_home>/garak if not already found in the environment
* visual_jailbreak downloaded artifacts are placed in <xdg_cache_home>/garak/resources
* generated data for beast/gcg/tap are placed in <xdg_cache_home>/garak/resources

Signed-off-by: Jeffrey Martin <jemartin@nvidia.com>

garak/_config.py

@@ -15,6 +15,11 @@
 import pathlib
 from typing import List
 import yaml
+from xdg_base_dirs import (
+    xdg_cache_home,
+    xdg_config_home,
+    xdg_data_home,
+)
 
 DICT_CONFIG_AFTER_LOAD = False
 
@@ -26,6 +31,7 @@
 run_params = "seed deprefix eval_threshold generations probe_tags interactive".split()
 plugins_params = "model_type model_name extended_detectors".split()
 reporting_params = "taxonomy report_prefix".split()
+project_dir = "garak"
 
 
 loaded = False
@@ -53,9 +59,17 @@ class TransientConfig(GarakSubConfig):
     args = None  # only access this when determining what was passed on CLI
     run_id = None
     basedir = pathlib.Path(__file__).parents[0]
+    config_dir = xdg_config_home() / project_dir
+    data_dir = xdg_data_home() / project_dir
+    cache_dir = xdg_cache_home() / project_dir
     starttime = None
     starttime_iso = None
 
+    # initialize the user home and cache paths if they do not exist
+    config_dir.mkdir(mode=0o740, parents=True, exist_ok=True)
+    data_dir.mkdir(mode=0o740, parents=True, exist_ok=True)
+    cache_dir.mkdir(mode=0o740, parents=True, exist_ok=True)
+
 
 transient = TransientConfig()
 
@@ -151,7 +165,7 @@ def load_config(
 
     settings_files = [str(transient.basedir / "resources" / "garak.core.yaml")]
 
-    fq_site_config_filename = str(transient.basedir / site_config_filename)
+    fq_site_config_filename = str(transient.config_dir / site_config_filename)
     if os.path.isfile(fq_site_config_filename):
         settings_files.append(fq_site_config_filename)
     else:

garak/command.py

@@ -8,8 +8,10 @@
 
 
 def start_logging():
+    from garak import _config
+
     logging.basicConfig(
-        filename="garak.log",
+        filename=_config.transient.data_dir / "garak.log",
         level=logging.DEBUG,
         format="%(asctime)s  %(levelname)s  %(message)s",
     )
@@ -32,6 +34,7 @@ def start_run():
     import os
     import uuid
 
+    from pathlib import Path
     from garak import _config
 
     logging.info("started at %s", _config.transient.starttime_iso)
@@ -41,19 +44,25 @@ def start_run():
             "⚠️ The current/default config is optimised for speed rather than thoroughness. Try e.g. --config full for a stronger test, or specify some probes."
         )
     _config.transient.run_id = str(uuid.uuid4())  # uuid1 is safe but leaks host info
+    # why is report prefix a condition of placing file in the report_dir?
+    report_path = Path(_config.reporting.report_dir)
+    if not report_path.is_absolute():
+        logging.debug("relative report dir provided")
+        report_path = _config.transient.data_dir / _config.reporting.report_dir
+    if not os.path.isdir(report_path):
+        try:
+            report_path.mkdir(mode=0o740, parents=True, exist_ok=True)
+        except PermissionError as e:
+            raise PermissionError(
+                f"Can't create logging directory {report_path}, quitting"
+            ) from e
+
+    filename = f"garak.{_config.transient.run_id}.report.jsonl"
     if not _config.reporting.report_prefix:
-        if not os.path.isdir(_config.reporting.report_dir):
-            try:
-                os.mkdir(_config.reporting.report_dir)
-            except PermissionError as e:
-                raise PermissionError(
-                    f"Can't create logging directory {_config.reporting.report_dir}, quitting"
-                ) from e
-        _config.transient.report_filename = f"{_config.reporting.report_dir}/garak.{_config.transient.run_id}.report.jsonl"
+        filename = f"garak.{_config.transient.run_id}.report.jsonl"
     else:
-        _config.transient.report_filename = (
-            _config.reporting.report_prefix + ".report.jsonl"
-        )
+        filename = _config.reporting.report_prefix + ".report.jsonl"
+    _config.transient.report_filename = str(report_path / filename)
     _config.transient.reportfile = open(
         _config.transient.report_filename, "w", buffering=1, encoding="utf-8"
     )

garak/evaluators/base.py

@@ -5,6 +5,7 @@
 
 import json
 import logging
+from pathlib import Path
 from typing import Iterable
 
 from colorama import Fore, Style
@@ -72,12 +73,19 @@ def evaluate(self, attempts: Iterable[garak.attempt.Attempt]) -> None:
                             hitlog_mode = (
                                 "w" if _config.transient.hitlogfile is None else "a"
                             )
+                            # why is report prefix a condition of placing file in the report_dir?
+                            hitlog_path = Path(_config.transient.report_filename).parent
                             if not _config.reporting.report_prefix:
-                                hitlog_filename = f"{_config.reporting.report_dir}/garak.{_config.transient.run_id}.hitlog.jsonl"
+                                hitlog_filename = (
+                                    hitlog_path
+                                    / f"garak.{_config.transient.run_id}.hitlog.jsonl"
+                                )  #
                             else:
                                 hitlog_filename = (
-                                    _config.reporting.report_prefix + ".hitlog.jsonl"
+                                    hitlog_path
+                                    / f"{_config.reporting.report_prefix}.hitlog.jsonl"
                                 )
+
                             logging.info("hit log in %s", hitlog_filename)
                             _config.transient.hitlogfile = open(
                                 hitlog_filename,

garak/interactive.py

@@ -233,7 +233,7 @@ def settings_ns_provider(self) -> argparse.Namespace:
 def banner():
     """Display a random banner"""
     with open(
-        _config.transient.basedir / "resources/banners", "r", encoding="utf-8"
+        _config.transient.basedir / "resources" / "banners", "r", encoding="utf-8"
     ) as bannerfile:
         banners = bannerfile.read().lstrip().split("\n\n")
         print(Fore.LIGHTRED_EX + f"\n{random.choice(banners)}\n" + Fore.RESET)

garak/probes/tap.py

@@ -37,6 +37,7 @@
 """
 
 import logging
+from pathlib import Path
 from typing import List
 
 import tqdm
@@ -68,7 +69,7 @@ class TAPCached(Probe):
 
     def __init__(
         self,
-        prompts_location: str = _config.transient.basedir
+        prompts_location: Path = _config.transient.basedir
         / "resources"
         / "tap"
         / "data"

garak/probes/visual_jailbreak.py

@@ -49,7 +49,7 @@ class FigStep(Probe):
 
     def _load_SafeBench(self):
         safebench_data_dir = (
-            _config.transient.basedir / "resources" / "visual_jailbreak" / "SafeBench"
+            _config.transient.cache_dir / "resources" / "visual_jailbreak" / "SafeBench"
         )
         if not os.path.exists(safebench_data_dir):
             # make the dir
@@ -94,7 +94,7 @@ def probe(self, generator):
         self.prompts = [
             {
                 "text": prompt["text"],
-                "image": str(_config.transient.basedir / prompt["image"]),
+                "image": str(_config.transient.cache_dir / prompt["image"]),
             }
             for prompt in self.prompts
         ]
@@ -125,7 +125,7 @@ def probe(self, generator):
         self.prompts = [
             {
                 "text": prompt["text"],
-                "image": str(_config.transient.basedir / prompt["image"]),
+                "image": str(_config.transient.cache_dir / prompt["image"]),
             }
             for prompt in self.prompts
         ]

garak/resources/autodan/autodan.py

@@ -31,6 +31,9 @@
 autodan_resource_data = (
     garak._config.transient.basedir / "resources" / "autodan" / "data"
 )
+cached_autodan_resource_data = (
+    garak._config.transient.cache_dir / "resources" / "autodan" / "data"
+)
 autodan_parser = argparse.ArgumentParser(description="AutoDAN config")
 autodan_parser.add_argument(
     "--num_steps", type=int, default=100, help="Number of steps to run generation"
@@ -88,9 +91,9 @@ def autodan_generate(
     mutation_generator_name: str = "gpt-3.5-turbo",
     mutation_generator_type: str = "openai",
     hierarchical: bool = False,
-    out_path: str = str(autodan_resource_data / "autodan_prompts.txt"),
-    init_prompt_path: str = str(autodan_resource_data / "autodan_init.txt"),
-    reference_path: str = str(autodan_resource_data / "prompt_group.pth"),
+    out_path: Path = cached_autodan_resource_data / "autodan_prompts.txt",
+    init_prompt_path: Path = autodan_resource_data / "autodan_init.txt",
+    reference_path: Path = autodan_resource_data / "prompt_group.pth",
     low_memory: bool = False,
     random_seed: int = None,
 ):
@@ -109,9 +112,9 @@ def autodan_generate(
         mutation_generator_name (str): Name of model to use as the mutation generator
         mutation_generator_type (str): Type of model to use as the mutation generator
         hierarchical (bool): Whether ot use hierarchical GA
-        out_path (str): Path to write generated AutoDAN string
-        init_prompt_path (str): Path to initial prompts
-        reference_path (str): Path to reference prompt tensors
+        out_path (Path): Path to write generated AutoDAN string
+        init_prompt_path (Path): Path to initial prompts
+        reference_path (Path): Path to reference prompt tensors
         low_memory (bool): Whether to use low memory
         random_seed (int): Random seed, if used.
 

garak/resources/autodan/genetic.py

@@ -2,33 +2,54 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import gc
+import nltk.downloader
 import numpy as np
 import torch
 import random
 import openai
-import os
 import re
 import nltk
 from nltk.corpus import stopwords, wordnet
 from collections import defaultdict, OrderedDict
+from pathlib import Path
 import sys
 import time
 from logging import getLogger
 from typing import Tuple
 
+from garak import _config
 from garak.resources.autodan.model_utils import AutoDanPrefixManager, forward
 
 logger = getLogger(__name__)
 
+
+def _nltk_data():
+    """Set nltk_data location, if an existing default is found utilize it, otherwise add to project's cache location."""
+    from nltk.downloader import Downloader
+
+    default_path = Path(Downloader().default_download_dir())
+    if not default_path.exists():
+        # if path not found then place in the user cache
+        # get env var for NLTK_DATA, fallback to create in cachedir / nltk_data
+        logger.debug("nltk_data location not found using project cache location")
+        _nltk_data_path.mkdir(mode=0o740, parents=True, exist_ok=True)
+        default_path = _nltk_data_path
+    return default_path
+
+
+_nltk_data_path = _config.transient.cache_dir / "nltk_data"
+nltk.data.path.append(str(_nltk_data_path))
+
 # TODO: Refactor into setup.py
 try:
     _ = stopwords.words("english")
     _ = nltk.word_tokenize("This is a normal English sentence")
     _ = wordnet.synsets("word")
 except LookupError as e:
-    nltk.download("stopwords")
-    nltk.download("punkt")
-    nltk.download("wordnet")
+    download_path = _nltk_data()
+    nltk.download("stopwords", download_dir=download_path)
+    nltk.download("punkt", download_dir=download_path)
+    nltk.download("wordnet", download_dir=download_path)
 
 
 # TODO: Could probably clean up the inputs here by using imports.

garak/resources/beast/beast_attack.py

@@ -2,6 +2,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import logging
+from pathlib import Path
 from typing import Union, Optional, Tuple
 
 import torch
@@ -17,7 +18,7 @@
     "A chat between a curious user and an artificial intelligence assistant. "
     "The assistant gives helpful, detailed, and polite answers to the user's questions."
 )
-beast_resource_data = garak._config.transient.basedir / "resources" / "beast" / "data"
+beast_resource_data = garak._config.transient.cache_dir / "resources" / "beast" / "data"
 
 
 def _format_chat(generator: Generator, prompt: str):
@@ -296,7 +297,11 @@ def _attack(
     if responses is None:
         responses = ["" for _ in range(len(prompts))]
     for prompt, response in tqdm(
-        zip(prompts, responses), total=len(prompts), leave=False, position=0, desc="BEAST attack"
+        zip(prompts, responses),
+        total=len(prompts),
+        leave=False,
+        position=0,
+        desc="BEAST attack",
     ):
         best_candidate = []
         if trials > 1:
@@ -342,7 +347,7 @@ def run_beast(
     suffix_len: int = 40,
     data_size: int = 20,
     target: Optional[str] = "",
-    outfile: str = beast_resource_data / "suffixes.txt",
+    outfile: Path = beast_resource_data / "suffixes.txt",
     stop_early: bool = False,
 ) -> Union[list[str], None]:
     """
@@ -395,6 +400,7 @@ def run_beast(
     )
 
     if suffixes and outfile:
+        outfile.parent.mkdir(mode=0o740, parents=True, exist_ok=True)
         with open(outfile, "a") as f:
             for suffix in suffixes:
                 f.write(f"{suffix}\n")

garak/resources/common.py

@@ -1,4 +1,5 @@
 import logging
+import shutil
 import urllib.error
 from pathlib import Path
 import pandas as pd
@@ -42,12 +43,21 @@
 
 
 def load_advbench(size: int = 0) -> pd.DataFrame:
-    advbench_path = (
+    advbench_base_path = (
         garak._config.transient.basedir
         / "resources"
         / "advbench"
         / "harmful_behaviors.csv"
     )
+    advbench_path = (
+        garak._config.transient.cache_dir
+        / "resources"
+        / "advbench"
+        / "harmful_behaviors.csv"
+    )
+    if advbench_base_path.is_file() and not advbench_path.is_file():
+        shutil.copy2(advbench_base_path, advbench_path)
+
     if not advbench_path.is_file():
         try:
             hb = "https://raw.githubusercontent.com/llm-attacks/llm-attacks/main/data/advbench/harmful_behaviors.csv"