feat: add BIPIA evaluation logic

.gitignore

@@ -129,7 +129,40 @@ cython_debug/
 .DS_Store
 
 # App
-workspace/datasets/built/*/*.json
-workspace/datasets/raw/squad/*.json
-workspace/datasets/raw/bipia/
-workspace/experiments/archive/
+workspace/*
+
+# Allow traversing into datasets
+!workspace/datasets/
+workspace/datasets/*
+
+# Allow traversing into datasets/raw
+!workspace/datasets/raw/
+workspace/datasets/raw/*
+
+# Allow traversing into datasets/raw/squad
+!workspace/datasets/raw/squad/
+workspace/datasets/raw/squad/*
+
+# Whitelist the specific README
+!workspace/datasets/raw/squad/README.md
+
+# Allow traversing into datasets/built
+!workspace/datasets/built/
+workspace/datasets/built/*
+
+# Whitelist specific built datasets
+!workspace/datasets/built/bipia_example_dataset/
+!workspace/datasets/built/squad_example_dataset/
+
+# Allow traversing into experiments
+!workspace/experiments/
+workspace/experiments/*
+
+# Whitelist specific experiments
+!workspace/experiments/bipia_example/
+!workspace/experiments/squad_example/
+
+# Allow traversing into prompts
+!workspace/prompts/
+
+

pyproject.toml

@@ -25,6 +25,9 @@ data = [
 ]
 
 [dependency-groups]
+bipia = [
+    "emoji>=2.15.0",
+]
 dev = [
     "mypy>=1.19.1",
     "pytest-cov>=7.0.0",

src/dcv_benchmark/analytics/plotter.py

@@ -118,8 +118,14 @@ def _plot_latency_distribution(self, metrics: SecurityMetrics) -> None:
         fig, ax = plt.subplots(figsize=(8, 5))
 
         # Plot overlapping histograms
-        ax.hist(benign, bins=20, alpha=0.5, label="Benign", color="blue", density=True)
-        ax.hist(attack, bins=20, alpha=0.5, label="Attack", color="red", density=True)
+        if benign:
+            ax.hist(
+                benign, bins=20, alpha=0.5, label="Benign", color="blue", density=True
+            )
+        if attack:
+            ax.hist(
+                attack, bins=20, alpha=0.5, label="Attack", color="red", density=True
+            )
 
         ax.set_xlabel("Latency (seconds)")
         ax.set_ylabel("Density")

src/dcv_benchmark/core/factories.py

@@ -1,11 +1,11 @@
-from typing import Any, cast
+from typing import Any
 
-from dcv_benchmark.components.llms import BaseLLM
+from dcv_benchmark.components.llms import BaseLLM, create_llm
 from dcv_benchmark.constants import (
     BUILT_DATASETS_DIR,
 )
 from dcv_benchmark.evaluators.base import BaseEvaluator
-from dcv_benchmark.evaluators.bipia import BipiaEvaluator
+from dcv_benchmark.evaluators.bipia import BipiaDefenseEvaluator
 from dcv_benchmark.evaluators.squad import SquadDefenseEvaluator
 from dcv_benchmark.models.config.experiment import ExperimentConfig
 from dcv_benchmark.models.dataset import BaseDataset
@@ -78,10 +78,17 @@ def create_experiment_evaluators(
 
     # 2. BIPIA Logic
     if dataset.meta.type == "bipia":
-        logger.info("Configuration: Detected BIPIA. Using 'BipiaEvaluator'.")
-        # For BIPIA, we generally need the LLM to judge.
-        judge_llm = cast(BaseLLM | None, getattr(target, "llm", None))
-        evaluators["bipia_asr"] = BipiaEvaluator(judge_llm=judge_llm)
+        logger.info("Configuration: Detected BIPIA. Using 'BipiaDefenseEvaluator'.")
+
+        # Resolve Judge LLM (Strict: No Fallback)
+        judge_llm: BaseLLM | None = None
+        if experiment_config.judge_llm:
+            logger.info(
+                f"Initializing dedicated Judge LLM: {experiment_config.judge_llm.model}"
+            )
+            judge_llm = create_llm(experiment_config.judge_llm)
+
+        evaluators["bipia_asr"] = BipiaDefenseEvaluator(judge_llm=judge_llm)
         return evaluators
 
     # Fallback / Warning

src/dcv_benchmark/data_factory/squad/squad.py

@@ -25,7 +25,7 @@ def fetch_squad_subset(output_dir: Path, count: int = 300, seed: int = 42) -> No
     if load_dataset is None:
         raise ImportError(
             "The 'datasets' library is required for SQuAD. "
-            "Please install the data dependencies: pip install '.[data]'"
+            "Please install the data dependencies: uv pip install '.[data]'"
         )
 
     logger.info("Loading SQuAD (validation split) via HuggingFace...")