Question on xgboost ranking:pairwise with ndcg eval_metric

Currently, I am trying to train xgboost on a ranking dataset. But I am seeing the following behaviour:

My parameters are as below:
params = {'booster': 'gbtree', 'colsample_bytree': 0.50, 'eta': 0.001, 
          'eval_metric': 'ndcg', 'gamma': 4, 'lambda': 10, 'max_depth': 8,
          'min_child_weight': 20.0, 'nthread': 4, 'objective': 'rank:pairwise', 'max_delta_step': 1
          , 'subsample': 0.9500000000000001, 'tree_method': 'exact'}
watchlist = [(train_pool, 'train'), (eval_pool, 'eval')]
progress = {}
model = xgb.train(
            params=params,
    num_boost_round=200,
    dtrain=train_pool,
    evals = watchlist,
    evals_result = progress,
    verbose_eval = True)