File size: 6,777 Bytes
4413aa8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3

import json
import os
import re
from pathlib import Path
from typing import List, Dict, Optional
from concurrent.futures import ThreadPoolExecutor, as_completed
from tqdm import tqdm
import numpy as np
from model_client import ModelClient


PROJECT_ROOT = Path(__file__).parent.parent

DEFAULT_DATA_VERSION = os.getenv("DATA_VERSION", "v1.1-paper")
DEFAULT_DATA_ROOT = Path(os.getenv("DATA_ROOT", PROJECT_ROOT / "data" / DEFAULT_DATA_VERSION)).expanduser()
DEFAULT_DEFINITION_FILE = DEFAULT_DATA_ROOT / "prompt" / "Definition.json"
DEFAULT_JUDGE_PROMPT_FILE = DEFAULT_DATA_ROOT / "prompt" / "reason_quality_judge.json"


class ReasonEvaluator:
    def __init__(
        self,
        judge_model: str = "gpt-5.2",
        temperature: float = 0.3,
        max_workers: int = 16,
        definition_file: Path = DEFAULT_DEFINITION_FILE,
        judge_prompt_file: Path = DEFAULT_JUDGE_PROMPT_FILE
    ):
        self.max_workers = max_workers
        self.definitions = self._load_json(definition_file)
        self.judge_prompt = self._load_judge_prompt(judge_prompt_file)
        self.client = ModelClient(model=judge_model, temperature=temperature)
    
    def _load_json(self, file_path: Path) -> Dict:
        with open(file_path, 'r', encoding='utf-8') as f:
            return json.load(f)
    
    def _load_judge_prompt(self, file_path: Path) -> str:
        data = self._load_json(file_path)
        if isinstance(data, str):
            return data
        return data.get("reason_quality_judge", data.get("prompt", ""))
    
    def _extract_reasoning(self, text: str) -> str:
        text = text or ""
        text = re.sub(r'<think>.*?</think>', '', text, flags=re.DOTALL)
        reason_match = re.search(r"<reason>(.*?)</reason>", text, flags=re.DOTALL | re.IGNORECASE)
        if reason_match:
            return reason_match.group(1).strip()

        text = re.sub(r"<score>.*?</score>", "", text, flags=re.DOTALL | re.IGNORECASE)
        text = re.sub(r'\$\$\$.*?\$\$\$', '', text)
        return text.strip()
    
    def _parse_score(self, response: str) -> Optional[float]:
        for pattern in [r'^(0\.0|0\.2|0\.4|0\.6|0\.8|1\.0)$', r'(0\.0|0\.2|0\.4|0\.6|0\.8|1\.0)']:
            match = re.search(pattern, response.strip(), re.IGNORECASE)
            if match:
                try:
                    score = float(match.group(1))
                    if score in [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]:
                        return score
                except:
                    continue
        return None
    
    def _format_input(self, dimension: str, definition: str, title: str, true_reasons: str, pred_output: str) -> str:
        pred_reasoning = self._extract_reasoning(pred_output)
        return f"""**Dimension**: {dimension}

**Dimension Definition**: {definition}

**Paper Title**: {title}

**Reference Reasoning**: {true_reasons}

**Predicted Output**: {pred_reasoning}

Please provide your quality score (0.0, 0.2, 0.4, 0.6, 0.8, or 1.0):"""
    
    def _judge_sample(self, idx: int, pred: Dict, dimension: str, definition: str) -> Dict:
        user_input = self._format_input(
            dimension, definition,
            pred.get('paper_title', ''),
            pred.get('true_reasons', ''),
            pred.get('pred_reasoning') or pred.get('pred_output', '')
        )
        
        try:
            response = self.client.call(user_text=user_input, system_prompt=self.judge_prompt)
            score = self._parse_score(response) or -1.0
            return {
                'index': idx, 'uid': pred.get('uid', ''), 'quality_score': score,
                'true_score': pred.get('true_score', 0), 'pred_score': pred.get('pred_score', 0)
            }
        except Exception as e:
            return {'index': idx, 'uid': pred.get('uid', ''), 'quality_score': -1.0, 'error': str(e)}
    
    def evaluate_reasons(self, predictions: List[Dict], dimension: str, show_progress: bool = True) -> Dict:
        dimension_key = dimension.capitalize()
        definition = self.definitions.get(dimension_key, "")
        if not definition:
            raise ValueError(f"Definition not found for '{dimension_key}'")
        
        results = []
        
        with ThreadPoolExecutor(max_workers=self.max_workers) as executor:
            futures = {
                executor.submit(self._judge_sample, i, p, dimension, definition): i
                for i, p in enumerate(predictions)
            }
            
            if show_progress:
                pbar = tqdm(total=len(predictions), desc=f"RQS {dimension}", ncols=80)
            
            for future in as_completed(futures):
                results.append(future.result())
                if show_progress:
                    pbar.update(1)
            
            if show_progress:
                pbar.close()
        
        results.sort(key=lambda x: x['index'])
        
        valid = [r['quality_score'] for r in results if r['quality_score'] >= 0]
        
        if not valid:
            stats = {
                'total_samples': len(results), 'valid_samples': 0, 'failed_samples': len(results),
                'mean_score': 0.0, 'median_score': 0.0, 'std_score': 0.0, 'score_distribution': {}
            }
        else:
            distribution = {}
            for s in [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]:
                count = sum(1 for v in valid if v == s)
                distribution[str(s)] = {'count': count, 'percentage': round(count / len(valid) * 100, 2)}
            
            stats = {
                'total_samples': len(results), 'valid_samples': len(valid),
                'failed_samples': len(results) - len(valid),
                'mean_score': round(float(np.mean(valid)), 4),
                'median_score': round(float(np.median(valid)), 4),
                'std_score': round(float(np.std(valid)), 4),
                'score_distribution': distribution
            }
        
        return {'dimension': dimension, 'scored_results': results, 'statistics': stats}


def evaluate_reasons_from_results(
    predictions: List[Dict],
    dimension: str,
    judge_model: str = "gpt-5.2",
    temperature: float = 0.3,
    max_workers: int = 16,
    show_progress: bool = True,
    definition_file: Path = DEFAULT_DEFINITION_FILE,
    judge_prompt_file: Path = DEFAULT_JUDGE_PROMPT_FILE
) -> Dict:
    evaluator = ReasonEvaluator(
        judge_model=judge_model,
        temperature=temperature,
        max_workers=max_workers,
        definition_file=definition_file,
        judge_prompt_file=judge_prompt_file
    )
    return evaluator.evaluate_reasons(predictions, dimension, show_progress)


def save_reason_evaluation(evaluation_result: Dict, output_dir: Path):
    return None, None