def test_bakeoff_dev_d3_9c():
global all_markables_dev
f,r,p = coref.eval_predictions('predictions/bakeoff-dev.preds',all_markables_dev)
assert f > .73
def test_bakeoff_vs_prof_d3_9():
global all_markables_te
# students can't run this
f,r,p = coref.eval_predictions('predictions/bakeoff-te.preds',all_markables_te)
assert f > .723
def test_test_acc_f1_d2_6():
f,r,p = coref.eval_predictions('predictions/rules-test.preds',all_markables_te);
assert_greater(f,.675)
assert_greater(r,.57)
assert_greater(p,.82)
def test_average_perceptron_d3_4b():
f,r,p = coref.eval_predictions('predictions/minimal-dev.preds',all_markables_dev);
assert(f>.66)
assert(r>.57)
assert(p>.78)
def test_bakeoff_vs_prof_d3_9():
global all_markables_te
# students can't run this
f,r,p = coref.eval_predictions('predictions/bakeoff-te.preds',all_markables_te)
assert f > .723
def test_dev_acc_f1_d2_6():
f,r,p = coref.eval_predictions('predictions/rules-dev.preds',all_markables_dev);
assert_greater(f,.68)
assert_greater(r,.58)
assert_greater(p,.81)
def test_average_perceptron_d3_4b():
f,r,p = coref.eval_predictions('predictions/minimal-dev.preds',all_markables_dev);
assert(f>.66)
assert(r>.57)
assert(p>.78)
def test_bakeoff_dev_d3_9c():
global all_markables_dev
f,r,p = coref.eval_predictions('predictions/bakeoff-dev.preds',all_markables_dev)
assert f > .73
def test_test_acc_f1_d2_6():
f,r,p = coref.eval_predictions('predictions/rules-test.preds',all_markables_te);
assert_greater(f,.675)
assert_greater(r,.57)
assert_greater(p,.82)
def test_dev_acc_f1_d2_6():
f,r,p = coref.eval_predictions('predictions/rules-dev.preds',all_markables_dev);
assert_greater(f,.68)
assert_greater(r,.58)
assert_greater(p,.81)
def test_test_acc_f1_d2_7():
f, r, p = coref.eval_predictions('predictions/rules-test.preds',
all_markables_te)
assert_greater(f, .499)
assert_greater(r, .381)
assert_greater(p, .723)
def test_dev_acc_f1_d2_7():
f, r, p = coref.eval_predictions('predictions/rules-dev.preds',
all_markables_dev)
assert_greater(f, .546)
assert_greater(r, .395)
assert_greater(p, .880)
def test_test_acc_f1_d2_7():
f, r, p = coref.eval_predictions('predictions/rules-test.preds', all_markables_te);
assert_greater(f, .499)
assert_greater(r, .381)
assert_greater(p, .723)
