Update README.md

README.md

@@ -1,7 +1,7 @@
 * A set of unstructured sparse bert-base-uncased models fine-tuned for SQuADv1.
 * Tensorflow models are created using ```TFAutoModelForQuestionAnswering.from_pretrained(..., from_pt=True)``` and ```model.save_pretrained(tf_pth)```.
 * Observed issue - loss in model translation, discrepancy observed in evaluation between pytorch and tensorflow models.
-* Table below is evaluated in HF's transformers v4.9.2.
+* Table below is evaluated in HF's transformers v4.9.2. Sparsity is normalized to dense layers in attention heads and FFNN.
 * Evaluation cli:
     ```bash
     python run_qa.py \