improved --SET-list and avise start-up loading time

avise/connectors/languagemodel/ollama.py

@@ -3,8 +3,6 @@
 import logging
 from typing import List
 
-import ollama
-
 from .base import BaseLMConnector, Message
 from ...registry import connector_registry
 from ...utils import ansi_colors
@@ -32,6 +30,9 @@ def __init__(
             config: Dictionary containing data from Connector configuration JSON.
             evaluation: Boolean flag indicating whether initializing a connector for the target model or the evaluation model
         """
+        # Imported here so that importing AVISE does not load the Ollama SDK
+        import ollama
+
         self.api_key = None
         if evaluation:
             self.model = config["eval_model"]["name"]

avise/connectors/languagemodel/openai.py

@@ -6,8 +6,6 @@
 import logging
 from typing import List
 
-from openai import OpenAI
-
 from .base import BaseLMConnector, Message
 from ...registry import connector_registry
 from ...utils import ansi_colors
@@ -133,6 +131,9 @@ def __init__(self, config: dict, evaluation: bool = False):
             else:
                 self.max_tokens = 512
 
+        # Imported here so that importing AVISE does not load the OpenAI SDK
+        from openai import OpenAI
+
         # Initialize OpenAI client
         try:
             client_kwargs = {"api_key": self.api_key}

avise/engine.py

@@ -255,11 +255,21 @@ def list_available(
         listings = []
         if sets:
             listings.append(
-                (set_registry, "avise.sets", "Available Security Evaluation Tests", "--SET-list")
+                (
+                    set_registry,
+                    "avise.sets",
+                    "Available Security Evaluation Tests",
+                    "--SET-list",
+                )
             )
         if connectors:
             listings.append(
-                (connector_registry, "avise.connectors", "Available Connectors", "--connector-list")
+                (
+                    connector_registry,
+                    "avise.connectors",
+                    "Available Connectors",
+                    "--connector-list",
+                )
             )
 
         found = True
@@ -296,7 +306,14 @@ def list_available(
 
         if reportformats:
             title = "Available Report Formats"
-            print("\n" + (f"{ansi_colors['bold']}{title}{ansi_colors['reset']}" if style.color else title))
+            print(
+                "\n"
+                + (
+                    f"{ansi_colors['bold']}{title}{ansi_colors['reset']}"
+                    if style.color
+                    else title
+                )
+            )
             for format in ReportFormat:
                 print(f"   {format.value}")
         return found

avise/models/adversarial_lm.py

@@ -8,17 +8,6 @@
 from typing import Optional
 from importlib.resources import files
 
-from transformers import (
-    Mistral3ForConditionalGeneration,
-    MistralCommonBackend,
-    AutoModelForCausalLM,
-    AutoTokenizer,
-    pipeline,
-)
-import torch
-from huggingface_hub import snapshot_download
-from mistral_common.exceptions import MistralCommonException
-
 logger = logging.getLogger(__name__)
 
 
@@ -41,6 +30,17 @@ def __init__(
         system_prompt: str = None,
         use_device: Optional[str] = "auto",
     ):
+        # Heavy ML libraries are imported here, not at module level, so that
+        # importing AVISE (e.g. for `avise --SET-list`) stays fast.
+        import torch
+        from transformers import (
+            Mistral3ForConditionalGeneration,
+            MistralCommonBackend,
+            AutoModelForCausalLM,
+            AutoTokenizer,
+        )
+        from mistral_common.exceptions import MistralCommonException
+
         logger.info("Loading Adversarial Language Model...")
         torch.cuda.empty_cache()
 
@@ -190,6 +190,8 @@ def generate_response(self, prompt, reasoning: bool = True) -> list:
         if "mistral" in self.model_name:
             response = self._mistral_text_generation(messages)
         else:
+            from transformers import pipeline
+
             model_pipeline = pipeline(
                 task="text-generation",
                 model=self.model,
@@ -224,6 +226,8 @@ def generate_response(self, prompt, reasoning: bool = True) -> list:
 
     def del_model(self):
         """Delete the model from GPU memory."""
+        import torch
+
         self.model.cpu()
         del self.model
         del self.tokenizer
@@ -270,6 +274,9 @@ def _model_download(
             model_path (str): Path where to save the model.
             model_name (str): Name of the Hugging Face model.
         """
+        from huggingface_hub import snapshot_download
+        from transformers import AutoModelForCausalLM, AutoTokenizer
+
         model_path = Path(files("avise").joinpath(model_path))
 
         # Check if path exists

avise/models/evaluation_lm.py

@@ -6,11 +6,6 @@
 import sys
 from importlib.resources import files
 
-import torch
-from transformers import Mistral3ForConditionalGeneration, MistralCommonBackend
-from huggingface_hub import snapshot_download
-from mistral_common.exceptions import MistralCommonException
-
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.DEBUG)
 
@@ -34,6 +29,12 @@ def __init__(
         system_prompt: str = None,
         use_device: str = "auto",
     ):
+        # Heavy ML libraries are imported here, not at module level, so that
+        # importing AVISE (e.g. for `avise --SET-list`) stays fast.
+        import torch
+        from transformers import Mistral3ForConditionalGeneration, MistralCommonBackend
+        from mistral_common.exceptions import MistralCommonException
+
         logger.info("Loading Evaluation Language Model...")
         torch.cuda.empty_cache()
         self.model = None
@@ -195,6 +196,8 @@ def _mistral_text_generation(self, messages: list) -> str:
     def del_model(self):
         """Delete the model from GPU memory."""
         if self.model:
+            import torch
+
             self.model.cpu()
             del self.model
             del self.tokenizer
@@ -213,6 +216,8 @@ def _model_download(
             model_path (str): Path where to save the model.
             model_name (str): Name of the Hugging Face model.
         """
+        from huggingface_hub import snapshot_download
+
         model_path = Path(files("avise").joinpath(model_path))
         # Check if path exists
         if not os.path.exists(model_path):

avise/pipelines/continuallearning/pipeline.py

@@ -17,6 +17,7 @@
 from typing import List, Dict, Any, Optional
 from datetime import datetime
 from math import sqrt
+from statistics import NormalDist
 
 from .schema import (
     ContinualLearningSETCase,
@@ -26,7 +27,6 @@
 )
 from ...connectors.continuallearning.base import BaseCLConnector
 
-from scipy.special import erfinv
 
 logger = logging.getLogger(__name__)
 
@@ -372,7 +372,11 @@ def _calculate_confidence_interval(
             return (0, 0.0, 0.0)
 
         p = passed / n
-        z = 1.96 if confidence_level == 0.95 else sqrt(2) * erfinv(confidence_level)
+        z = (
+            1.96
+            if confidence_level == 0.95
+            else NormalDist().inv_cdf((1 + confidence_level) / 2)
+        )
 
         denominator = 1 + (z**2 / n)
         center = (p + (z**2 / (2 * n))) / denominator

avise/pipelines/languagemodel/pipeline.py

@@ -11,12 +11,12 @@
 from typing import List, Dict, Any, Optional
 from datetime import datetime
 from math import sqrt
+from statistics import NormalDist
 
 from .schema import LanguageModelSETCase, OutputData, EvaluationResult, ReportData
 from ...connectors.languagemodel.base import BaseLMConnector
 from ...models import EvaluationLanguageModel
 
-from scipy.special import erfinv
 
 logger = logging.getLogger(__name__)
 
@@ -339,7 +339,11 @@ def _calculate_confidence_interval(
         p = passed / n
 
         # Z-score for confidence level (1.96 for 95% CI)
-        z = 1.96 if confidence_level == 0.95 else sqrt(2) * erfinv(confidence_level)
+        z = (
+            1.96
+            if confidence_level == 0.95
+            else NormalDist().inv_cdf((1 + confidence_level) / 2)
+        )
 
         # Wilson score interval
         denominator = 1 + (z**2 / n)

avise/sets/continuallearning/backdoor.py

@@ -242,8 +242,8 @@ class Backdoor(BaseSETPipeline):
 
     name = "Backdoor"
     description = (
-        "Injects a backdoor trigger into a Continual Learning model's training data and "
-        "evaluates whether the model learns to associate the trigger with a target label."
+        "Tests whether poisoning a continual learning model's training data with a "
+        "backdoor trigger makes it classify triggered inputs as an attacker-chosen label."
     )
 
     def __init__(self):

avise/sets/languagemodel/multi_turn/context_test.py

@@ -30,7 +30,10 @@ class ContextTest(BaseSETPipeline):
     """Context Test SET."""
 
     name = "Context test"
-    description = "A simple test for multi-turn conversations where the conversation history is passed to the target"
+    description = (
+        "Tests whether the target model keeps track of the conversation context "
+        "across multiple turns, with the full history sent in every request."
+    )
 
     def __init__(self):
         super().__init__()