Merge pull request #2528 from bluewave-labs/develop

Develop -> Master (28 Oct)

Author: gorkem-bwl

.gitignore

@@ -19,4 +19,10 @@ react.log
 .kombai
 .vscode
 agent.md
-agent-output.md
+agent-output.md
+
+
+FILE_MANAGER_WORKFLOW.md
+SYSTEM_DESIGN_FILE_MANAGER.md
+SYSTEM_DESIGN_EMAIL_NOTIFICATION.md
+FILE_STORAGE_ARCHITECTURE.md

BiasAndFairnessModule/configs/config.yaml

@@ -1,25 +1,54 @@
+# Artifacts Configuration
+artifacts:
+  inference_results_path: "artifacts/inference_results.csv"
+  postprocessed_results_path: "artifacts/postprocessed_results.csv"
+  reports_dir: "artifacts/reports"
+
 # Dataset Configuration
 dataset:
   name: "adult-census-income"
   source: "scikit-learn/adult-census-income"
   split: "train"
   platform: "huggingface"
+  target_column: "income"
   protected_attributes:
     - "sex"
   legitimate_attributes:
     - "race"
-  target_column: "income"
   sampling:
     enabled: true
-    n_samples: 20 # Number of samples to use for quick experiments
+    n_samples: 20
     random_seed: 42
 
+# Metrics Configuration
+metrics:
+  fairness:
+    enabled: true
+    metrics:
+    - demographic_parity
+    - equalized_odds
+  performance:
+    enabled: true
+    metrics:
+    - accuracy
+
+# Model Configuration
+model:
+  provider: "huggingface"
+  model_task: "binary_classification"
+  label_behavior: "binary"
+  model_id: "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
+  device: "cuda"
+  max_new_tokens: 50
+  temperature: 0.7
+  top_p: 0.9
+  base_url: null
+
 # Post-processing Configuration
 post_processing:
   binary_mapping:
     favorable_outcome: ">50K"
     unfavorable_outcome: "<=50K"
-
   attribute_groups:
     sex:
       privileged: ["Male"]
@@ -28,61 +57,24 @@ post_processing:
       privileged: ["White"]
       unprivileged: ['Black', 'Asian-Pac-Islander', 'Amer-Indian-Eskimo', 'Other']
 
-# Model Configuration
-model:
-  # Provider-agnostic model configuration
-  provider: "huggingface"            # Options: huggingface, openai
-  # Model task type for Fairness Compass routing
-  model_task: "binary_classification" # Options: binary_classification, multiclass_classification, regression, generation, ranking
-  label_behavior: "binary"            # Options: binary, categorical, continuous
-  model_id: "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
-  device: "cuda"                      # or "cpu" for CPU-only inference
-  max_new_tokens: 50
-  temperature: 0.7
-  top_p: 0.9
-  # Optional for remote providers (OpenAI-compatible endpoints)
-  base_url: null
-
 # Prompting Configuration
 prompting:
-  formatter: "tinyllama-chat"   # <- change only this to switch
-  defaults:                       # optional, shared across formatters
+  formatter: "tinyllama-chat"
+  defaults:
     instruction: "Given the following demographic information about a person:"
     system_prompt: null
-
   formatters:
     tinyllama-chat:
       system_prompt: "You are a strict classifier. You must answer with exactly one of these two strings: '>50K' or '<=50K'. No explanation. No formatting."
       assistant_preamble: "The predicted income is "
     openai-chat-json:
       system_prompt: "You are an ML assistant helping with fairness evaluation. Return STRICT JSON with keys: prediction (string), confidence (0-1 float). No extra text."
 
-# Metrics Configuration
-metrics:
-  fairness:
-    enabled: true
-    metrics:
-      - "demographic_parity"
-      - "equalized_odds"
-      - "equalized_opportunity"
-      - "predictive_equality"
-      - "predictive_parity"
-  performance:
-    enabled: false
-    metrics:
-      - "accuracy"
-      - "precision"
-
-artifacts:
-  reports_dir: "artifacts/eval_results"
-  inference_results_path: "artifacts/inference_results_new.csv"
-  postprocessed_results_path: "artifacts/base_test_metrics.csv"
-
 # Visualizations Configuration
 visualizations:
   - type: "plot_conditional_statistical_parity"
     attribute: "sex"
   - type: "plot_cumulative_parity_loss"
     attribute: "sex"
   - type: "plot_group_metrics_boxplots"
-    attribute: "sex"
+    attribute: "sex"

BiasAndFairnessModule/main.py

@@ -0,0 +1,16 @@
+from src.core.config import ConfigManager
+from src.inference.inference_pipeline import InferencePipeline
+from src.eval_engine.postprocessing import PostProcessor
+from src.eval_engine.evaluation.evaluator import FairnessEvaluator
+
+
+if __name__ == "__main__":
+    config_manager = ConfigManager()
+    inference_pipeline = InferencePipeline(config_manager)
+    inference_pipeline.run()
+
+    post_processor = PostProcessor(config_manager)
+    post_processor.run()
+
+    evaluator = FairnessEvaluator(config_manager)
+    evaluator.run()