Update README.md

README.md

@@ -58,23 +58,26 @@ SFT LoRA обучение было выполнено на **двух NVIDIA A10
 - Fused AdamW  
 - Liger Kernel (swiglu, fused linear xentropy)  
   
-**GPU hours**: ~384 of NVIDIA A100  
+**GPU hours**: ~384h of NVIDIA A100  
 **GPU mem**:  
 - Stage 1: 50-55GB of VRAM (both GPUs)
 - Stage 2: 79GB of VRAM (both GPUs)
   
 ### Training configuration / Конфигурация обучения
-**The model was trained using MyLLM framework:**  
+The model was trained using MyLLM framework:  
+  
 --== [MyLLM](https://github.com/Raumberg/myllm) ==--  
+  
 **Model training / Обучение модели**  
+  
 The model was trained utilizing 2 stages:
 - Stage 1:
   - Datasets: GrandMaster, LoRA: rank=128, alpha=256
 - Stage 2:
   - Datasets: Kolmogorov-3, Russian Code, LoRA: rank=256, alpha=256
-
-**All configs are available in MyLLM repository.**  
-
+  
+All configs are available in MyLLM repository.  
+  
 ### Using the model / Как запустить?
 
 ```python