File size: 8,584 Bytes
926675f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
202
203
204
205
206
207
208
209
210
211
212
import re
import string
from collections import Counter

import numpy as np
import pandas as pd
import tqdm
from langchain.evaluation.qa import QAEvalChain
from langchain.llms import OpenAI

from algos.PWS import PWS_Base, PWS_Extra
from algos.notool import CoT, IO
from algos.react import ReactBase


def normalize_answer(s):
    def remove_articles(text):
        return re.sub(r"\b(a|an|the)\b", " ", text)

    def white_space_fix(text):
        return " ".join(text.split())

    def remove_punc(text):
        exclude = set(string.punctuation)
        return "".join(ch for ch in text if ch not in exclude)

    def lower(text):
        return text.lower()

    return white_space_fix(remove_articles(remove_punc(lower(s))))


def f1_score(prediction, ground_truth):
    normalized_prediction = normalize_answer(prediction)
    normalized_ground_truth = normalize_answer(ground_truth)

    if normalized_prediction in ['yes', 'no', 'noanswer'] and normalized_prediction != normalized_ground_truth:
        return 0
    if normalized_ground_truth in ['yes', 'no', 'noanswer'] and normalized_prediction != normalized_ground_truth:
        return 0

    prediction_tokens = normalized_prediction.split()
    ground_truth_tokens = normalized_ground_truth.split()
    common = Counter(prediction_tokens) & Counter(ground_truth_tokens)
    num_same = sum(common.values())
    if num_same == 0:
        return 0
    precision = 1.0 * num_same / len(prediction_tokens)
    recall = 1.0 * num_same / len(ground_truth_tokens)
    f1 = (2 * precision * recall) / (precision + recall)
    return f1


def llm_accuracy_score(query, prediction, ground_truth):
    data = [{
        'query': query,
        'answer': ground_truth,
    }]
    pred = [{
        'query': query,
        'answer': ground_truth,
        'result': prediction,
    }]
    eval_chain = QAEvalChain.from_llm(OpenAI(temperature=0))
    graded_outputs = eval_chain.evaluate(data, pred)
    return 1 if graded_outputs[0]['text'].strip() == 'CORRECT' else 0


class Evaluator:
    def __init__(self, task, dataset, algo, maxtry=3):
        assert task in ["hotpot_qa", "trivia_qa", "gsm8k", "physics_question", "disfl_qa",
                        "sports_understanding", "strategy_qa", "sotu_qa"]
        assert isinstance(dataset, pd.DataFrame)
        assert isinstance(algo, (PWS_Base, PWS_Extra, ReactBase, IO, CoT))

        self.task = task
        self.dataset = dataset
        self.algo = algo
        self.maxtry = maxtry
        self.failed_response = self._failed_response()
        self.eval_data = self._initialize_eval_dict()

    def run(self):
        print("\n******************* Start Evaluation *******************\n")
        if self.task in ["hotpot_qa", "sotu_qa"]:
            for i in tqdm.tqdm(range(len(self.dataset))):
                question = self.dataset["question"][i]
                label = self.dataset["answer"][i]
                for _ in range(self.maxtry):
                    try:
                        response = self.algo.run(question)
                        break
                    except:
                        response = self.failed_response
                self._update_eval_dict(question, label, response)

        elif self.task == "fever":
            for i in tqdm.tqdm(range(len(self.dataset))):
                question = self.dataset["claim"][i]
                label = self.dataset["label"][i]
                for _ in range(self.maxtry):
                    try:
                        response = self.algo.run(question)
                        break
                    except:
                        response = self.failed_response
                self._update_eval_dict(question, label, response)
        elif self.task == "trivia_qa":
            for i in tqdm.tqdm(range(len(self.dataset))):
                question = self.dataset["question"][i]
                label = self.dataset["answer"][i]["value"]
                for _ in range(self.maxtry):
                    try:
                        response = self.algo.run(question)
                        break
                    except:
                        response = self.failed_response
                self._update_eval_dict(question, label, response)
        elif self.task == "gsm8k":
            for i in tqdm.tqdm(range(len(self.dataset))):
                question = self.dataset["question"][i]
                label = self.dataset["answer"][i].split("#### ")[1]
                for _ in range(self.maxtry):
                    try:
                        response = self.algo.run(question)
                        break
                    except:
                        response = self.failed_response
                self._update_eval_dict(question, label, response)
        elif self.task in ["physics_question", "sports_understanding", "strategy_qa"]:
            for i in tqdm.tqdm(range(len(self.dataset))):
                question = self.dataset["input"][i]
                label = self.dataset["target"][i]
                for _ in range(self.maxtry):
                    try:
                        response = self.algo.run(question)
                        break
                    except:
                        response = self.failed_response
                self._update_eval_dict(question, label, response)
        else:
            raise NotImplementedError

        return self._get_avg_results(), self.eval_data

    def _initialize_eval_dict(self):
        data = {}
        for d in ["label", "preds", "em", "f1", "acc", "wall_time", "total_tokens", "total_cost", "steps", "token_cost",
                  "tool_cost", "planner_log", "solver_log"]:
            data[d] = []
        return data

    def _update_eval_dict(self, question, label, response):
        pred = self._parse_prediction(response["output"])
        self.eval_data["label"] += [label]
        self.eval_data["preds"] += [pred]
        self.eval_data["em"] += [self.get_metrics(question, label, pred)["em"]]
        self.eval_data["f1"] += [self.get_metrics(question, label, pred)["f1"]]
        self.eval_data["acc"] += [self.get_metrics(question, label, pred)["acc"]]
        self.eval_data["wall_time"] += [response["wall_time"]]
        self.eval_data["total_tokens"] += [response["total_tokens"]]
        self.eval_data["total_cost"] += [response["total_cost"]]
        self.eval_data["steps"] += [response["steps"]]
        self.eval_data["token_cost"] += [response["token_cost"]]
        self.eval_data["tool_cost"] += [response["tool_cost"]]

        if "planner_log" in response:
            self.eval_data["planner_log"] += [response["planner_log"]]
        if "solver_log" in response:
            self.eval_data["solver_log"] += [response["solver_log"]]

    def _get_avg_results(self):
        result = {}
        result["avg_em"] = np.nanmean(self.eval_data["em"])
        result["avg_f1"] = np.nanmean(self.eval_data["f1"])
        result["avg_acc"] = np.nanmean(self.eval_data["acc"])
        result["avg_wall_time"] = np.nanmean(self.eval_data["wall_time"])
        result["avg_total_tokens"] = np.nanmean(self.eval_data["total_tokens"])
        result["avg_total_cost"] = np.nanmean(self.eval_data["total_cost"])
        result["avg_steps"] = np.nanmean(self.eval_data["steps"])
        result["avg_token_cost"] = np.nanmean(self.eval_data["token_cost"])
        result["avg_tool_cost"] = np.nanmean(self.eval_data["tool_cost"])
        return result

    def get_metrics(self, query, label, pred):
        if pred is None:
            return {'em': 0, 'f1': 0}
        norm_label = normalize_answer(label)
        norm_pred = normalize_answer(pred)
        em = (norm_pred == norm_label)
        f1 = f1_score(norm_pred, norm_label)
        acc = llm_accuracy_score(query, pred, label)
        return {'em': em, 'f1': f1, 'acc': acc}

    def _parse_prediction(self, output):
        if isinstance(self.algo, IO):
            return str(output).strip("\n")
        elif isinstance(self.algo, CoT):
            return str(output).split("\n")[-1].replace("Answer:", "")
        elif isinstance(self.algo, ReactBase):
            return str(output).strip("\n")
        elif isinstance(self.algo, PWS_Base):
            return str(output).strip("\n")
        elif isinstance(self.algo, PWS_Extra):
            return str(output).strip("\n")

    def _failed_response(self):
        resposne = {}
        for key in ["input", "output", "wall_time", "total_tokens", "total_cost", "steps", "token_cost", "tool_cost"]:
            resposne[key] = np.nan
        return resposne