fix dataset card link, clarify bundled checkpoint in training section

Author: phanii9

docs/model_card.md

@@ -100,10 +100,12 @@ python -m meow.inference \
 ### Training Data
 
 20,000 synthetic (input, output) samples across 15 categories.
-See the [dataset card](dataset_card.md) for full details.
+See the [dataset card](https://huggingface.co/datasets/hunt3rx99/meowllm-miso) for full details.
 
 ### Training Procedure
 
+The recommended training regime (for a fresh checkpoint):
+
 - **Optimizer**: AdamW (β₁=0.9, β₂=0.95, weight_decay=0.1)
 - **Learning rate**: 3e-4 peak, linear warmup + cosine decay
 - **Warmup**: 200 steps
@@ -115,6 +117,10 @@ See the [dataset card](dataset_card.md) for full details.
 - **Hardware**: single T4 GPU (free Colab tier sufficient)
 - **Wall-clock time**: ~20 minutes
 
+> **Note:** The bundled `best.pt` was trained on CPU for 2000 steps
+> (batch size 32, ~4 minutes). See the Measured numbers section below
+> for its real eval results. A full GPU run will produce higher numbers.
+
 ### Training hyperparameters (MeowConfig)
 
 ```python