v2: updated model card

README.md

@@ -1,20 +1,18 @@
 ---
 library_name: sklearn
 tags:
-  - tabular-classification
+  - fairrelay
   - logistics
   - xgboost
   - sklearn
+  - tabular-classification
   - fairness
-  - gini-index
-  - fairrelay
 datasets:
   - Cainiao-AI/LaDe-D
-  - electricsheepafrica/africa-synth-retail-and-ecommerce-last-mile-delivery-data-nigeria
 license: mit
 ---
 
-# FairRelay — Fairness Classification Model (ACCEPT vs REOPTIMIZE)
+# FairRelay — Fairness Classification Model (ACCEPT vs REOPTIMIZE) (v2)
 
 Part of the **[FairRelay](https://github.com/MUTHUKUMARAN-K-1/FairRelay)** AI logistics platform.
 
@@ -22,30 +20,46 @@ Part of the **[FairRelay](https://github.com/MUTHUKUMARAN-K-1/FairRelay)** AI lo
 
 Fairness Classification Model (ACCEPT vs REOPTIMIZE)
 
-**Type**: XGBClassifier Pipeline  
-**Framework**: scikit-learn Pipeline + XGBoost  
+**Version**: v2 — Retrained with realistic, harder data to prevent overfitting and improve real-world robustness.
+
+**Type**: XGBoost Pipeline (StandardScaler + XGBoost)
 **Task**: Classification
 
+### v2 Improvements Over v1
+
+- **Hidden confounders**: Weather, traffic, building access affect ground truth but aren't in features
+- **Heteroscedastic noise**: Harder cases have more unpredictable outcomes
+- **Non-linear interactions**: Weight × stairs, packages × rain compound effects
+- **Measurement error**: Features have ±5-15% sensor/estimation noise
+- **Boundary ambiguity**: Near-threshold cases have noisy labels (simulating dispatcher disagreement)
+- **Diverse distributions**: Normal, skewed, bimodal, heavy-tail effort patterns
+
 ## Performance
 
-- **Accuracy**: 0.9992
-- **F1 Score**: 0.9994
+- **Accuracy**: 0.9000
+- **F1 Score**: 0.9369
+- **Precision**: 0.9266
+- **Recall**: 0.9474
+- **Train-Test Gap**: 0.0211
+- **CV F1 (5-fold)**: 0.9401 ± 0.0012
 
 
 ## Input Features
 
 | Feature | Importance |
 |---------|-----------|
-| `num_drivers` | 0.0280 |
-| `avg_effort` | 0.0012 |
-| `std_dev` | 0.2209 |
-| `max_gap` | 0.7071 |
-| `gini_index` | 0.0096 |
-| `min_effort` | 0.0022 |
-| `max_effort` | 0.0012 |
-| `outlier_count` | 0.0179 |
-| `pct_above_avg` | 0.0013 |
-| `effort_cv` | 0.0106 |
+| `num_drivers` | 0.0255 |
+| `avg_effort` | 0.0151 |
+| `std_dev` | 0.1706 |
+| `max_gap` | 0.5543 |
+| `gini_index` | 0.0585 |
+| `min_effort` | 0.0152 |
+| `max_effort` | 0.0209 |
+| `outlier_count` | 0.0605 |
+| `pct_above_avg` | 0.0138 |
+| `effort_cv` | 0.0334 |
+| `skewness` | 0.0145 |
+| `kurtosis` | 0.0176 |
 
 
 ## Usage
@@ -53,33 +67,19 @@ Fairness Classification Model (ACCEPT vs REOPTIMIZE)
 ```python
 from skops import io as sio
 from huggingface_hub import hf_hub_download
+import numpy as np
 
-# Download and load
 model_path = hf_hub_download(repo_id="muthuk1/fairrelay-fairness-classifier", filename="model.skops")
 untrusted = sio.get_untrusted_types(file=model_path)
 model = sio.load(model_path, trusted=untrusted)
 
-# Predict
-import numpy as np
-features = np.array([[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]])
 prediction = model.predict(features)
 ```
 
-## Training Data
-
-- **Cainiao-AI/LaDe-D**: Real last-mile delivery data from Shanghai (KDD 2023)
-- **Africa Synth Last-Mile**: Synthetic Nigerian delivery data
-- **FairRelay Synthetic**: Physics-informed synthetic data calibrated to FairRelay's deterministic formulas
-
 ## Part of FairRelay
 
-FairRelay is an AI-powered logistics platform for fair load consolidation and dispatch:
-- 🚚 5-agent load consolidation pipeline (KMeans + OR-Tools CP-SAT)
-- ⚖️ 8-agent fair dispatch pipeline (Gini optimization)
-- 📊 XGBoost effort prediction + Thompson Sampling bandit
-- 🌱 EV-aware routing with battery constraints
-
-Built for **LogisticsNow Hackathon 2026** — Challenge #5: AI Load Consolidation
+FairRelay is an AI-powered logistics platform for fair load consolidation and dispatch.
+Built for **LogisticsNow Hackathon 2026** — Challenge #5: AI Load Consolidation.
 
 ## License