@@ -24,25 +24,25 @@ def setUpClass(cls):
2424
2525 def test_basic_correct_response (self ):
2626 """Test if semantically similar responses are marked correct."""
27- response = "Density; Velocity; Viscosity; Length"
28- answer = "Density; Velocity; Viscosity; Length"
27+ response = "Density, Velocity, Viscosity, Length"
28+ answer = "Density, Velocity, Viscosity, Length"
2929 result = evaluation_function (response , answer , self .param )
3030
3131 self .assertTrue (result .get ("is_correct" ))
3232
3333 def test_basic_incorrect_response (self ):
3434 """Test if semantically different responses are marked incorrect."""
35- response = "Mass; Speed; Friction; Force"
36- answer = "Density; Velocity; Viscosity; Length"
35+ response = "Mass, Speed, Friction, Force"
36+ answer = "Density, Velocity, Viscosity, Length"
3737 result = evaluation_function (response , answer , self .param )
3838
3939
4040 self .assertFalse (result .get ("is_correct" ))
4141
4242 def test_partial_match (self ):
4343 """Test if a response too short is marked incorrect."""
44- response = "Density; Velocity; Viscosity"
45- answer = "Density; Velocity; Viscosity; Length"
44+ response = "Density, Velocity, Viscosity"
45+ answer = "Density, Velocity, Viscosity, Length"
4646 result = evaluation_function (response , answer , self .param )
4747 self .assertFalse (result .get ("is_correct" ))
4848
@@ -57,25 +57,25 @@ def test_synonyms_match(self):
5757
5858 def test_exact_match_requirement (self ):
5959 """Test enforcing exact match on keystrings."""
60- response = "density; velocity; viscosity; length"
61- answer = "Density; Velocity; Viscosity; Length"
60+ response = "density, velocity, viscosity, length"
61+ answer = "Density, Velocity, Viscosity, Length"
6262
6363 result = evaluation_function (response , answer , self .param )
6464
6565 self .assertTrue (result .get ("is_correct" ))
6666
6767 def test_should_not_contain (self ):
6868 """Test if a response with a prohibited keyword fails."""
69- response = "density; velocity; viscosity; length; direction"
70- answer = "Density; Velocity; Viscosity; Length"
69+ response = "density, velocity, viscosity, length, direction"
70+ answer = "Density, Velocity, Viscosity, Length"
7171
7272 result = evaluation_function (response , answer , self .param )
7373
7474 self .assertFalse (result .get ("is_correct" ))
7575
7676 def test_negation_handling (self ):
7777 """Test how the model handles negation."""
78- response = "not light blue; dark blue"
78+ response = "not light blue, dark blue"
7979 answer = "light blue"
8080
8181 result = evaluation_function (response , answer , self .param )
@@ -86,7 +86,7 @@ def test_negation_handling(self):
8686 def test_short_response (self ):
8787 """Test how the model handles negation."""
8888 response = "yellow"
89- answer = "yellow; blue"
89+ answer = "yellow, blue"
9090
9191 result = evaluation_function (response , answer , self .param )
9292
@@ -95,8 +95,8 @@ def test_short_response(self):
9595
9696 def test_performance (self ):
9797 """Ensure that processing time is reasonable."""
98- response = "Density; Velocity; Viscosity; Length"
99- answer = "Density; Velocity; Viscosity; Length"
98+ response = "Density, Velocity, Viscosity, Length"
99+ answer = "Density, Velocity, Viscosity, Length"
100100
101101 result = evaluation_function (response , answer , self .param )
102102
0 commit comments