LSTM for small dataset

I test LSTM with a 2-category dataset with each dataset 800 data, I trained 100 epochs and try to make the model overfit and I expect to see a 100% accuracy for the training data prediction. However I failed and LSTM learn nothing, the taining accuracy is always 50% and the loss isn’t decreasing at all.
(When I try with each dataset 80k data with the same model, LSTM learn something and the prediction is around 70%.)
Anyone can tell me why? Thank you!