File size: 10,813 Bytes
0a026c0
 
dca2e30
0a026c0
 
 
 
 
 
 
 
 
1bad9c0
bb4bda1
 
0a026c0
 
d3bca1f
0a026c0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0955149
 
0a026c0
 
0955149
fd72b9a
0a026c0
 
 
595f661
da04c4f
2316f10
0a026c0
 
 
 
 
 
 
 
 
da04c4f
 
0a026c0
 
 
d3bca1f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea070cc
 
 
491fd45
 
 
 
 
ea070cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0eb1a66
ea070cc
 
0eb1a66
0a026c0
d3bca1f
0a026c0
d3bca1f
 
 
 
 
 
 
 
 
 
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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
import json

from util.assistants import GPTAgent
import json_repair

class evaluator:
    def __init__(self, model_name='GPT4-turbo'):
        self.model = GPTAgent(model_name)

    def validate_scores(self, scores):
        required_keys = ["Factually Correct", "Useful", "Context Specific", "User Specific", "Provides Pluralism"]
        for key in required_keys:
            if key not in scores or not isinstance(scores[key], (int, float)) or not (-1 <= scores[key] <= 1):
                return {"Factually Correct": -1,"Useful": -1,"Context Specific": -1,"User Specific":-1,"Provides Pluralism":-1}

        return scores

    def evaluate_single(self, question,explanation):

        evaluation_prompt = f"""You are provided with a user's question and the corresponding explanation generated by 
        an AI model. Your task is to evaluate the explanation based on the following five principles. Each principle 
        should be scored on a scale from 0 to 1, where 0 indicates that the principle is not met at all, 
        and 1 indicates that the principle is fully satisfied.
        
        Question: 
        {question}
        
        Provided Explanation: 
        {explanation}
        
        Evaluation Criteria:

        Factually Correct:
        Definition: The explanation must be accurate and relevant to the question and the subject matter.
        Score: (0-1) How factually correct is the explanation? Consider the accuracy of the details provided and their relevance to the question.
        
        Useful:
        Definition: The explanation should enable the user to understand the answer better and should facilitate further reasoning or decision-making.
        Score: (0-1) How useful is the explanation in helping the user understand the answer and make informed decisions?
        
        Context Specific:
        Definition: The explanation should be relevant to the specific context or scenario implied by the question.
        Score: (0-1) How well does the explanation address the specific context or scenario of the question?
        
        User Specific:
        Definition: The explanation should cater to the knowledge level and interests of the user, assuming typical or specified user characteristics.
        Score: (0-1) How well does the explanation cater to the needs and knowledge level of the intended user?
        
        Provides Pluralism:
        Definition: The explanation should offer or accommodate multiple viewpoints or interpretations, allowing the user to explore various perspectives.
        Score: (0-1) How well does the explanation provide or support multiple perspectives?
    
        After evaluating the provided question and explanation based on the five principles, please format your scores in a JSON dictionary. Directly provide me with the json without any additional text.
 
        Example JSON format:
        
        Answer:{{"Factually Correct": 0.9,"Useful": 0.85,"Context Specific": 0.8,"User Specific": 0.75,"Provides Pluralism": 0.7}}
        
        Answer:
        """

        response = self.model.invoke(evaluation_prompt,temperature=0, max_tokens=500).strip()
        #response = """{{"Factually Correct": 0.9,"Useful": 0.85,"Context Specific": 0.8,"User Specific": 0.75,"Provides Pluralism": 0.7}}"""
        print(response)
        try:
            scores = json.loads(response)
        except json.JSONDecodeError:
            # Attempt to repair the JSON if decoding fails
            repaired_json = json_repair.repair_json(response, skip_json_loads=True, return_objects=False)
            try:
                scores = json.loads(repaired_json)
            except json.JSONDecodeError:
                print("Failed to decode JSON response even after repair attempt. Skipping this batch.")
                return {"Factually Correct": -1,"Useful": -1,"Context Specific": -1,"User Specific":-1,"Provides Pluralism":-1}


        return self.validate_scores(scores)

    def format_conversation(self, conversation):
        formatted_conversation = "\n".join(
            f"{exchange['role'].capitalize()}: {exchange['content']}" for exchange in conversation
        )
        return formatted_conversation

    def evaluate_conversation(self, conversation, context):
        formatted_conversation = self.format_conversation(conversation)
        evaluation_prompt = f"""
        You are provided with a conversation between a user and a chatbot and the context about them. Your task is to evaluate the chatbot explanation in the conversation based on the following five principles. Each principle should be scored on a scale from 0 to 1, where 0 indicates that the principle is not met at all, and 1 indicates that the principle is fully satisfied.

        Conversation:
        {formatted_conversation}

        Context:
        {context}

        Evaluation Criteria:

        Factually Correct:
        Definition: The explanation must be accurate and relevant to the question and the subject matter.
        Score: (0-1) How factually correct is the explanation? Consider the accuracy of the details provided and their relevance to the question.

        Useful:
        Definition: The explanation should enable the user to understand the answer better and should facilitate further reasoning or decision-making.
        Score: (0-1) How useful is the explanation in helping the user understand the answer and make informed decisions?

        Context Specific:
        Definition: The explanation should be relevant to the specific context or scenario implied by the question.
        Score: (0-1) How well does the explanation address the specific context or scenario of the question?

        User Specific:
        Definition: The explanation should cater to the knowledge level and interests of the user, assuming typical or specified user characteristics.
        Score: (0-1) How well does the explanation cater to the needs and knowledge level of the intended user?

        Provides Pluralism:
        Definition: The explanation should offer or accommodate multiple viewpoints or interpretations, allowing the user to explore various perspectives.
        Score: (0-1) How well does the explanation provide or support multiple perspectives?

        After evaluating the provided conversation based on the context and five principles, please format your scores in a JSON dictionary. Directly provide me with the json without any additional text.

        Example JSON format:

        Answer: {{"Factually Correct": 0.9, "Useful": 0.85, "Context Specific": 0.8, "User Specific": 0.75, "Provides Pluralism": 0.7}}

        Answer:
        """

        print(evaluation_prompt)

        response = self.model.invoke(evaluation_prompt, temperature=0, max_tokens=500).strip()
        try:
            scores = json.loads(response)
        except json.JSONDecodeError:
            repaired_json = json_repair.repair_json(response, skip_json_loads=True, return_objects=False)
            try:
                scores = json.loads(repaired_json)
            except json.JSONDecodeError:
                print("Failed to decode JSON response even after repair attempt. Skipping this batch.")
                return {key: -1 for key in ["Factually Correct", "Useful", "Context Specific", "User Specific", "Provides Pluralism"]}

        return self.validate_scores(scores)

def write_evaluation_commentary(scores):
    evaluation_details = []
    for principle, score in scores.items():

        if score == -1:
            evaluation_details.append({'Principle': principle, 'Score': score, 'Commentary': 'Failed to evaluate the explanation.'})
            continue

        if principle == "Factually Correct":
            if score >= 0.8:
                comment = "Excellent accuracy! The information is precise and directly relevant to the question."
            elif score >= 0.5:
                comment = "Moderately accurate, but some details may not be completely correct or are somewhat irrelevant."
            else:
                comment = "The explanation contains significant inaccuracies or irrelevant information."
        elif principle == "Useful":
            if score >= 0.8:
                comment = "Highly useful! The explanation clearly enhances understanding and aids in further reasoning or decision-making."
            elif score >= 0.5:
                comment = "Somewhat useful, though it could be more insightful or practical in aiding understanding."
            else:
                comment = "The explanation does little to help understand or apply the information provided."
        elif principle == "Context Specific":
            if score >= 0.8:
                comment = "Perfectly tailored to the context of the question, addressing the specific scenario effectively."
            elif score >= 0.5:
                comment = "Generally addresses the context, but may miss specific details or nuances relevant to the question."
            else:
                comment = "Fails to address the context of the question, lacking relevance or specificity."
        elif principle == "User Specific":
            if score >= 0.8:
                comment = "The explanation is well-adapted to the user's knowledge level and interests, demonstrating thoughtfulness."
            elif score >= 0.5:
                comment = "Moderately considerate of the user's knowledge level, but could be more tailored."
            else:
                comment = "Does not consider the user's background or interests, potentially leading to confusion or disinterest."
        elif principle == "Provides Pluralism":
            if score >= 0.8:
                comment = "Provides an excellent range of perspectives or interpretations, fostering a comprehensive understanding."
            elif score >= 0.5:
                comment = "Offers some alternative perspectives, but more could be provided to enrich understanding."
            else:
                comment = "Lacks diversity in viewpoints, limiting the depth of exploration into the topic."

        evaluation_details.append({'Principle': principle, 'Score': score, 'Commentary': comment})
    return evaluation_details

if __name__ == '__main__':

    eval = evaluator()
    conversation = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Who won the world series in 2020?"},
        {"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."},
        {"role": "user", "content": "Where was it played?"}
    ]
    context = "general user, user_background is sports enthusiast"
    results = eval.evaluate_conversation(conversation, context)
    print(results)
    print(write_evaluation_commentary(results))