-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathPrompts.txt
176 lines (135 loc) · 8.5 KB
/
Prompts.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
DM:
Task Description:
This task involves evaluating the output of a Natural Language Processing (NLP) model to determine if it constitutes a hallucination. A hallucination in this context means that the model's output contains information not supported by the provided semantic reference. The task will use the following inputs:
Source (src): This is the context in which a word is used. The word to be defined may or may not be enclosed within special tokens <define> and </define>.
Target (tgt): This is the intended, correct definition of the word, typically sourced from a reference like Wiktionary.
Hypothesis (hyp): This is the actual output produced by the NLP model - the definition it generated for the given word.
ALWAYS ASSUME BOTH THE TARGET AND SOURCE FOR SEMANTIC REFERENCE FOR EVALUATING THE HYPOTHESIS
The task is to assess the hypothesis against the indicated reference (source AND target) to determine if it constitutes a hallucination. The output will be the probability of the hypothesis being a hallucination, represented as a value in the range [0.0, 1.0]. This probability is derived from the consensus of five human annotators, with each increment of 0.2 in the probability representing one annotator's opinion.
Examples:
Input:
'hyp': 'Having a relationship ; connected .'
'src': 'Electric and magnetic forces are closely related .'
'tgt': 'Standing in relation or connection .'
Output:
p(Hallucination): 0.8
Input:
'hyp': 'To negotiate a price or terms of a transaction .'
'src': 'They had to bargain for a few minutes to get a decent price for the rug .'
'tgt': 'To make a bargain ; to make a deal or contract for the exchange of property or services ; to negotiate'
Output:
p(Hallucination): 0.0
Input:
'hyp': 'A religious service in which the sabbath is observed .'
'src': 'Tefillin are not used on Sabbaths and holidays , so unless you attend a minyan on weekday mornings or grew up in a home where men prayed daily , this practice may be unfamiliar to you .'
'tgt': 'A Jewish prayer service .'
Output:
p(Hallucination): 1.0
Input:
'hyp': '(slang, derogatory) A contemptible person.'
'src': 'And that <define> jiboney </define> across the hall . He makes life worse than it is . Where he gets his money for booze , who knows ?'
'tgt': 'A low-level tough or thug, a mean man.'
Output:
p(Hallucination): 0.6
Input:
'hyp': '(informal) A test to determine whether something has a certain vibe.'
'src': "Usually we 'll listen to a track . It just depends . Sometimes they have a funny premise or a certain vibe and we 'll make something oriented towards that premise or vibe . If it passes the <define> vibe check </define> , we kick things off ."
'tgt': '(slang) An impromptu attempt to ascertain mood, opinions, or attributes.',
Output:
p(Hallucination): 0.4
Input:
'hyp': '(transitive) To make tacky.'
'src': 'to <define> tackify </define> rubber'
'tgt': '(transitive) To make (a material) tacky.'
Output:
p(Hallucination): 0.2
PG:
Task Description:
This task involves evaluating the output of a Natural Language Processing (NLP) model to determine if it constitutes a hallucination. A hallucination in this context means that the model's output contains information not supported by the provided semantic reference. The task will use the following inputs:
Source (src): input passed to be parphrased to NLP model
Target (tgt): This is the intended, correct paraphrased "gold" text
Hypothesis (hyp): This is the actual output produced by the NLP model - the paraphrase it generated for the given word.
ALWAYS ASSUME BOTH THE TARGET AND SOURCE FOR SEMANTIC REFERENCE FOR EVALUATING THE HYPOTHESIS
The task is to assess the hypothesis against the indicated reference (source AND target) to determine if it constitutes a hallucination. The output will be the probability of the hypothesis being a hallucination, represented as a value in the range [0.0, 1.0]. This probability is derived from the consensus of five human annotators, with each increment of 0.2 in the probability representing one annotator's opinion.
Examples:
Input:
'src': 'Let me out of here.'
'tgt': 'Someone let me out.'
'hyp': 'Let me leave.'
Output:
p(Hallucination): 0.0
Input:
'src': 'I have a permit.'
'tgt': 'Uh, I'm validated.
'hyp': 'I have a permit to carry it.'
Output:
p(Hallucination): 1.0
Input:
'hyp': "You're gonna leave the talking to me alone."
'src': 'Leave the talking to me.'
'tgt': 'Let me do all the talking.
Output:
p(Hallucination): 0.4
Input:
'hyp': "We don't have the money to risk it, all right?"
'src': "We can't afford to risk it."
'tgt': "We can't risk that."
Output:
p(Hallucination): 0.6
Input:
'hyp': 'It is not impossible.'
'src': 'Nothing is impossible.'
'tgt': 'There's nothing that can't be done.'
Output:
p(Hallucination): 0.2
Input:
'hyp': 'A number between five and eight.'
'src': 'A five, six, seven, eight.'
'tgt': 'And 5, 6, 7, 8.',
Output:
p(Hallucination): 0.8
MT:
Task Description:
This task involves evaluating the output of a Natural Language Processing (NLP) model to determine if it constitutes a hallucination. A hallucination in this context means that the model's output contains information not supported by the provided semantic reference. The task will use the following inputs:
Source (src): input passed to be parphrased to NLP model
Target (tgt): This is the intended, correct translation of the sentence or the "gold" text
Hypothesis (hyp): This is the actual output produced by the NLP model - the translation it generated for the given sentence.
ALWAYS ASSUME BOTH THE TARGET AND SOURCE FOR SEMANTIC REFERENCE FOR EVALUATING THE HYPOTHESIS
The task is to assess the hypothesis against the indicated reference (source AND target) to determine if it constitutes a hallucination. The output will be the probability of the hypothesis being a hallucination, represented as a value in the range [0.0, 1.0]. This probability is derived from the consensus of five human annotators, with each increment of 0.2 in the probability representing one annotator's opinion.
Examples:
Input:
'hyp': 'He is expected to run for president in 2016.'
'src': 'Akuganiziridwa kuti azayima nawo pa chisankho cha purezidenti mucha chaka cha 2016.'
'tgt': 'He is speculated to make a run for president in 2016.'
Output:
p(Hallucination): 0.0
Input:
'src': 'Andre biorytme-baserte alternativ involverer å drikke mye væske (særlig vann eller te, kjente vanndrivende væsker) før søvn, som vil tvinge en til å stå opp for å urinere'
'tgt': 'Other biorhythm-based options involve drinking lots of fluid (particularly water or tea, a known diuretic) prior to sleep, forcing one to get up to urinate.'
'hyp': 'Another biorhythm-based alternative involves drinking plenty of fluid (such as water or tea, known as water-driven fluid) before bedtime, which will force one to stand up to urinate.'
Output:
p(Hallucination): 0.2
Input:
'hyp': 'Pinagsapata and Martelly were the speakers of the nine-member Provisional Electoral Council (CEP).'
'src': 'Pinagsapata ni Martelly idi kalman ti baro a Provisional Electoral Council (CEP) nga addaan siam a miembro.'
'tgt': 'Martelly swore in a new Provisional Electoral Council (CEP) of nine members yesterday.'
Output:
p(Hallucination): 0.4
Input:
"hyp": "Six detainees, including children and the elderly, were released as Filipino photographers."
'src': "Pillillu īkā vuddulu sā 6 mādi baṇḍīlu Filipino bōtōgrābarlu lāgā vīduḍalu cheyāru."
"tgt": "Six hostages, including the children and elderly, were released early, as were the Filipino photographers."
Output:
p(Hallucination): 0.6
Input:
"hyp": "The bill could be amended, a unanimous state-level agreement was needed, but the state-level central government acted so lightly that its representatives were often not hierarchical."
"src": "Lai panti varētu tikt grozīti, bija nepieciešama visu štatu vienbalsīga piekrišana, bet štati pret centrālo valdību izturējās tik vieglprātīgi, ka to pārstāvji bieži vien nebija ieradušies."
"tgt": "The Articles required unanimous consent from all the states before they could be amended and states took the central government so lightly that their representatives were often absent."
Output:
p(Hallucination): 0.8
Input:
"hyp": "The oven is full of food and drinks."
"src": "فرنو ضغ نا فرمد بضن غتين كرو فرمد ضا غد بى صسنغن"
"tgt": "The scenes are displayed on the pyramids and the different pyramids are lit up."
Output:
p(Hallucination): 1.0