Update README.md

README.md

@@ -155,10 +155,10 @@ outputs = model.generate(
 ### LoRA Configuration
 ```python
 {
-    "r": 8,                    # LoRA rank
-    "lora_alpha": 16,          # LoRA scaling factor
-    "lora_dropout": 0.05,      # Dropout probability
-    "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"],
+    "r": 8,
+    "lora_alpha": 16,
+    "lora_dropout": 0.05,
+    "target_modules": ["q_proj", "v_proj"],
     "task_type": "CAUSAL_LM"
 }
 ```
@@ -180,20 +180,18 @@ outputs = model.generate(
 learning_rate = 2e-4
 warmup_steps = 50
 max_steps = 500
-per_device_train_batch_size = 8
-gradient_accumulation_steps = 128
+per_device_train_batch_size = 16
+gradient_accumulation_steps = 4
 effective_batch_size = 1024
 
 # Optimization
 optimizer = "adamw_torch_xla"
 lr_scheduler = "cosine"
 weight_decay = 0.01
-max_grad_norm = 1.0
 
 # Model Settings
 sequence_length = 256
 precision = "bfloat16"
-gradient_checkpointing = True
 ```
 
 ### Training Infrastructure
@@ -239,15 +237,6 @@ The model was evaluated on the complete HumanEval benchmark (164 programming pro
 
 This demonstrates that the educational fine-tuning maintains strong algorithmic correctness while improving code clarity and documentation.
 
-### Sample Performance by Category
-
-| Category | Base Model | Fine-tuned | Delta |
-|----------|-----------|------------|-------|
-| String Manipulation | 68% | 65% | -3% |
-| Data Structures | 67% | 64% | -3% |
-| Algorithms | 66% | 63% | -3% |
-| Math/Logic | 64% | 65% | +1% |
-
 ---
 
 ## 🎓 Use Cases