fix: disable eval (no test split in dataset)

training/configs/ornith-35b-lora.yaml

@@ -50,10 +50,8 @@ mixed_precision: bf16
 # Using accelerate device_map for DISTRIBUTED loading
 # No DeepSpeed - model already quantized
 
-# Evaluation
+# Evaluation (disable - no test split in dataset)
-eval_strategy: steps
+eval_strategy: no
-eval_steps: 100
-eval_accumulation_steps: 10
 
 # Gradient Checkpointing
 gradient_checkpointing: true