Update README.md

README.md

@@ -72,7 +72,7 @@ The model uses a Persian tokenizer named “بلبل زبان” available on Hu
 
 - **Training Loss:** Reduced from 10.3 to 0.005 over 60 epochs.
 - **Validation Loss:** Reduced from 10.1 to 0.010.
-- **Test Accuracy:** 66% (exact match required).
+- **Test Accuracy:** 79% (exact match required).
 - **Perplexity:** 1.01
 
 ## Inference