These results show reasonable and relevant results. The decline from 100% to 75% lead to the biggest decrease in precision by over 6%, while the decrease from 75% to as low as 20% reduces the MAP by approximately 5.6%, even though the number of samples were nearly quartered. Then, the decrease from 20% to 5% reduces MAP by approximately 4%. Overall, the decrease in the upper percentage range seems larger for both classes than the decrease in the lower range, while the results between 75% to 20% show a small difference for the large range span.
Considering the tensorboard output for the 10% validation suggests optimal training time is about 120k steps, further training decreases the MAP.