Update README.md

README.md

@@ -34,13 +34,13 @@ model-index:
           split: test
         metrics:
           - type: accuracy
-            value: 0.9652
+            value: 0.9784
             name: Accuracy
           - type: f1
-            value: 0.9101
+            value: 0.9412
             name: F1 (Toxic)
           - type: f1
-            value: 0.9728
+            value: 0.9845
             name: F1 (Safe)
 ---
 
@@ -57,22 +57,23 @@ Hayā is the core classifier in a larger **defense-in-depth content moderation s
 
 ### Key Highlights
 
-- 🎯 **96.52% accuracy** on a held-out test set of ~96K sentences
+- 🎯 **97.84% accuracy** on a held-out test set of ~100K sentences
 - 🌍 **Every major Arabic dialect**: Egyptian, Levantine, Gulf, Maghrebi, Iraqi, Sudanese, MSA
-- 📊 Trained on **~770K examples** from 44 curated public Arabic datasets
+- 📊 Trained on **nearly 1 million examples (997K)** from 51 curated public Arabic datasets
 - ⚡ Optimized for real-time inference via serverless API (Modal + FastAPI)
+- 🧠 **V2 Update:** Enhanced with continued fine-tuning on 34K hard edge-cases to conquer implicit hate and subtle toxicity.
 
 ---
 
 ## Performance
 
-Evaluated on a **held-out test set of ~96,000 sentences** (stratified, zero data leakage):
+Evaluated on a **held-out test set of 99,759 sentences** (stratified, zero data leakage):
 
 | Metric | Score |
 |--------|-------|
-| **Accuracy** | **96.52%** |
-| **F1 (Toxic)** | **91.01%** |
-| **F1 (Safe)** | **97.28%** |
+| **Accuracy** | **97.84%** |
+| **F1 (Toxic)** | **94.12%** |
+| **F1 (Safe)** | **98.45%** |
 
 > **Note on the numbers:** Manual error analysis showed the model frequently *outperformed the original human annotations* — many counted "errors" were actually mislabels in the source datasets. Real-world performance on correctly-labeled data is therefore higher than the raw scores suggest.
 
@@ -162,9 +163,9 @@ for text, result in zip(texts, results):
 
 ### Training Data
 - **Dataset**: [youssefreda9/HAYAA](https://huggingface.co/datasets/youssefreda9/HAYAA)
-- **Training examples**: ~770K
-- **Validation examples**: ~96K
-- **Sources**: 44 curated public Arabic hate-speech and abuse datasets
+- **Training examples**: ~798K
+- **Validation examples**: ~100K
+- **Sources**: 51 curated public Arabic hate-speech and abuse datasets
 - **Dialects**: All major Arabic dialects + MSA
 
 ### Hyperparameters
@@ -172,9 +173,9 @@ for text, result in zip(texts, results):
 | Parameter | Value |
 |-----------|-------|
 | Max sequence length | 128 |
-| Epochs | 5 (with early stopping, patience=2) |
+| Training Strategy | 2-Stage (4 epochs base corpus + 3 epochs hard edge-cases) |
 | Batch size | 16 (effective: 32 with gradient accumulation) |
-| Learning rate | 2e-5 |
+| Learning rate | 2e-5 (Stage 1), 5e-6 (Stage 2) |
 | Warmup ratio | 0.1 |
 | Weight decay | 0.01 |
 | FP16 | ✅ |
@@ -234,7 +235,7 @@ The full pipeline ships as a **Chrome Extension (Manifest V3)** with PIN-protect
 ## Limitations & Biases
 
 - **Platform bias**: Training data is predominantly from Twitter; performance may vary on other platforms (forums, messaging apps, gaming chat).
-- **Annotation noise**: Despite extensive cleaning, some label noise from the 44 original source datasets may persist.
+- **Annotation noise**: Despite extensive cleaning, some label noise from the 51 original source datasets may persist.
 - **Dialect imbalance**: Egyptian and MSA are better represented than Sudanese or Yemeni dialects.
 - **Max length**: Sequences longer than 128 tokens are truncated, which may affect classification of very long texts.
 - **Evolving language**: Arabic internet slang and evasion tactics evolve constantly; periodic retraining is recommended.