File size: 6,923 Bytes
14e4843
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
from typing import Union, List

from lm_eval.api.task import ConfigurableTask
from lm_eval.api.instance import Instance
# from lm_eval.api.registry import register_task
from lm_eval.api.metrics import mean

from src.backend.envs import DEVICE

import spacy
from selfcheckgpt.modeling_selfcheck import SelfCheckMQAG, SelfCheckNLI, SelfCheckBERTScore, SelfCheckNgram


# @register_task("selfcheckgpt")
class SelfCheckGPT(ConfigurableTask):
    VERSION = 0.0
    DATASET_PATH = "potsawee/wiki_bio_gpt3_hallucination"
    DATASET_NAME = None
    OUTPUT_TYPE = 'generate_until'

    def __init__(self):
        super().__init__(config={'metadata': {'version': self.VERSION}})
        # these end tokens are hard coded because of the current limitaion of the llm-eval.
        self.generation_kwargs = {"until": ["\n\n", "<unk>", "<|im_end|>", "</s>", "<|endoftext|>"], "max_length": 512}
        self.generation_kwargs_sampling_number = 5  # the number of sampling for self-consistence
        self.generation_kwargs_sampling = {"temperature": 0.99, "do_sample": True, "until": ["\n\n", "<unk>", "<|im_end|>", "</s>"], "max_length": 512}

        self.selfcheckgpt_type = os.environ.get('SELFCHECKGPTTYPE', 'SelfCheckNLI')
        self.selfcheckgpt_device = os.environ.get('SELFCHECKGPTDEVICE', DEVICE)
        self.selfcheckgpt_nlp = spacy.load("en_core_web_sm")

        if self.selfcheckgpt_type == 'SelfCheckNgram':
            self.selfcheckgpt = SelfCheckNgram(n=1)
        elif self.selfcheckgpt_type == 'SelfCheckBERTScore':
            self.selfcheckgpt = SelfCheckBERTScore(rescale_with_baseline=True)
        elif self.selfcheckgpt_type == 'SelfCheckMQAG':
            self.selfcheckgpt = SelfCheckMQAG(device=self.selfcheckgpt_device)
        elif self.selfcheckgpt_type == 'SelfCheckNLI':
            self.selfcheckgpt = SelfCheckNLI(device=self.selfcheckgpt_device)
        self.SelfCheckNLI_error_cnt = 0

    def has_training_docs(self):
        return False

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
        return False

    def validation_docs(self):
        return self.dataset["evaluation"]

    def doc_to_text(self, doc):
        if not hasattr(self, 'selfcheckgpt_nlp'):
            self.selfcheckgpt_nlp = spacy.load("en_core_web_sm")

        sentences = [x.text.strip() for x in self.selfcheckgpt_nlp(doc['wiki_bio_text']).sents]
        if len(sentences) < 2:
            raise ValueError("This wikipedia passage is too short for self-consistency check: {sentences}")
            # disscussed with Potsawee

        doc_text = f"Please generate a Wikipedia passage that consists of at least two sentences, starting with the following sentence: {sentences[0]}\n"
        return doc_text

    def doc_to_target(self, doc):
        answer = doc['wiki_bio_text']
        return answer

    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> Union[List[Instance], Instance]:
        arguments = (ctx, self.generation_kwargs)
        request_list = [
            Instance(request_type='generate_until', doc=doc, arguments=arguments, idx=0, **kwargs),
        ]
        sampling_arguments = (ctx, self.generation_kwargs_sampling)
        request_list.extend([
            Instance(request_type='generate_until', doc=doc, arguments=sampling_arguments, idx=idx, **kwargs)
            for idx in range(1, self.generation_kwargs_sampling_number+1)
            ]
        )
        return request_list

    def process_results(self, doc, results):
        response_temperature_0 = results[0]
        other_responses = results[1:]
        passage = self.doc_to_target(doc)

        sentences = self.selfcheckgpt_nlp(response_temperature_0)
        sentences = [sent.text.strip() for sent in sentences.sents]
        if self.selfcheckgpt_type == 'SelfCheckNgram':
            selfcheckgpt_scores = self.selfcheckgpt.predict(sentences=sentences, passage=response_temperature_0, sampled_passages=other_responses)
            return {
                'avg-selfcheckgpt': selfcheckgpt_scores['doc_level']['avg_neg_logprob'],
                'max-selfcheckgpt': selfcheckgpt_scores['doc_level']['avg_max_neg_logprob']
            }

        elif self.selfcheckgpt_type == 'SelfCheckBERTScore':
            selfcheckgpt_scores = self.selfcheckgpt.predict(sentences=sentences, sampled_passages=other_responses)
        elif self.selfcheckgpt_type == 'SelfCheckMQAG':
            selfcheckgpt_scores = self.selfcheckgpt.predict(
                sentences=sentences,
                passage=response_temperature_0,
                sampled_passages=other_responses,
                num_questions_per_sent=5,          # number of questions to be drawn
                scoring_method='bayes_with_alpha', # options = 'counting', 'bayes', 'bayes_with_alpha'
                beta1=0.8, beta2=0.8)            # additional params depending on scoring_method
        elif self.selfcheckgpt_type == 'SelfCheckNLI':
            selfcheckgpt_scores = self.selfcheckgpt.predict(sentences=sentences, sampled_passages=other_responses)

            if len(selfcheckgpt_scores) < 2:
                # at least two sentences
                self.SelfCheckNLI_error_cnt += 1
                result = {
                    'avg-selfcheckgpt': 0.0,
                    'max-selfcheckgpt': 0.0
                }

            else:
                threshold = 0.7 # https://huggingface.co/blog/dhuynh95/automatic-hallucination-detection
                # passage is hallucianted if one sentence is hallucinated. It's very strict.
                selfcheckgpt_scores_max = 0.0 if max(selfcheckgpt_scores) > threshold else 1.0
                # passage is hallucianted if average score of all sentences is hallucinated.
                selfcheckgpt_scores_avg = 0.0 if sum(selfcheckgpt_scores) / len(selfcheckgpt_scores) > threshold else 1.0
                result = {'avg-selfcheckgpt': selfcheckgpt_scores_avg, 'max-selfcheckgpt': selfcheckgpt_scores_max}

            return result

        selfcheckgpt_scores_avg = sum(selfcheckgpt_scores) / len(selfcheckgpt_scores) if len(selfcheckgpt_scores) > 0 else 0
        selfcheckgpt_scores_max = max(selfcheckgpt_scores)

        return {'avg-selfcheckgpt': selfcheckgpt_scores_avg, 'max-selfcheckgpt': selfcheckgpt_scores_max}

    def aggregation(self):
        """
        :returns: {str: [float] -> float}
            A dictionary where keys are the names of submetrics and values are
            functions that aggregate a list of metrics
        """
        return {k: mean for k in ["avg-selfcheckgpt", "max-selfcheckgpt"]}

    def higher_is_better(self):
        """
        :returns: {str: bool}
            A dictionary where keys are the names of submetrics and values are
            whether a higher value of the submetric is better
        """
        return {k: True for k in ["avg-selfcheckgpt", "max-selfcheckgpt"]}