Datasets:

ArXiv:
File size: 10,217 Bytes
9f3bc09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
import json
import random
from typing import List, Dict, Any
import numpy as np

class DatasetAugmenter:
    """Augment and enhance training dataset quality"""
    
    def __init__(self, base_dataset: str, annotations: str = None):
        self.base_data = self.load_json(base_dataset)
        self.annotations = self.load_json(annotations) if annotations else {}
        
    def load_json(self, file_path: str) -> Dict:
        """Load JSON data"""
        with open(file_path, 'r') as f:
            return json.load(f)
    
    def augment_with_variations(self, trajectory: List[Dict]) -> List[List[Dict]]:
        """Create variations of a trajectory"""
        variations = []
        
        # 1. Early stopping variation
        if len(trajectory) > 3:
            early_stop = trajectory[:random.randint(2, len(trajectory)-1)]
            variations.append(early_stop)
        
        # 2. Different tool selection
        tool_varied = []
        alternative_tools = {
            "Color Binding": ["Color", "Overall Consistency"],
            "Shape Binding": ["Object Class", "Multiple Objects"],
            "Subject Consistency": ["Appearance Style", "Overall Consistency"],
            "Motion Smoothness": ["Dynamic Degree", "Temporal Style"]
        }
        
        for step in trajectory:
            step_copy = step.copy()
            if step.get("tool") in alternative_tools:
                step_copy["tool"] = random.choice(alternative_tools[step["tool"]])
            tool_varied.append(step_copy)
        variations.append(tool_varied)
        
        # 3. Reordered exploration
        if len(trajectory) > 2:
            reordered = trajectory.copy()
            # Swap middle steps
            if len(reordered) > 3:
                idx1, idx2 = random.sample(range(1, len(reordered)-1), 2)
                reordered[idx1], reordered[idx2] = reordered[idx2], reordered[idx1]
            variations.append(reordered)
        
        return variations
    
    def add_negative_examples(self, good_trajectory: List[Dict]) -> List[Dict]:
        """Create negative examples from good trajectories"""
        negative_examples = []
        
        # 1. Redundant exploration (same tool/aspect multiple times)
        if len(good_trajectory) > 2:
            redundant = good_trajectory.copy()
            # Duplicate a random step
            dup_idx = random.randint(0, len(good_trajectory)-2)
            redundant.insert(dup_idx + 1, good_trajectory[dup_idx])
            negative_examples.append({
                "trajectory": redundant,
                "issue": "redundant_exploration",
                "quality": "poor"
            })
        
        # 2. Premature stopping
        if len(good_trajectory) > 3:
            premature = good_trajectory[:2]
            negative_examples.append({
                "trajectory": premature,
                "issue": "premature_stopping",
                "quality": "poor"
            })
        
        # 3. Wrong tool selection
        wrong_tool = []
        for step in good_trajectory:
            step_copy = step.copy()
            if random.random() < 0.3:  # 30% chance of wrong tool
                all_tools = ["Color Binding", "Shape Binding", "Texture Binding", 
                           "Subject Consistency", "Motion Smoothness", "Object Class"]
                step_copy["tool"] = random.choice(all_tools)
            wrong_tool.append(step_copy)
        negative_examples.append({
            "trajectory": wrong_tool,
            "issue": "inappropriate_tools",
            "quality": "poor"
        })
        
        return negative_examples
    
    def enrich_with_reasoning(self, trajectory: List[Dict]) -> List[Dict]:
        """Add detailed reasoning to trajectory steps"""
        enriched = []
        
        reasoning_templates = {
            "explore": [
                "Based on the previous results showing {observation}, we need to explore {aspect} to {goal}",
                "The model performed {performance} on {previous_aspect}, so testing {current_aspect} will help determine {insight}",
                "To fully answer the user's question about {topic}, examining {aspect} with {tool} is essential"
            ],
            "summarize": [
                "After {n} exploration steps, we have sufficient evidence that {conclusion}",
                "The consistent pattern across {aspects} indicates {finding}, providing a complete answer",
                "Further exploration would be redundant as we've established {key_insight}"
            ]
        }
        
        for i, step in enumerate(trajectory):
            step_copy = step.copy()
            
            # Add enriched reasoning
            if step.get("decision_type") == "explore":
                template = random.choice(reasoning_templates["explore"])
                step_copy["enriched_thought"] = template.format(
                    observation="strong performance" if i == 0 else "mixed results",
                    aspect=step.get("sub_aspect", "this aspect"),
                    goal="understand the model's capabilities",
                    performance="well" if random.random() > 0.5 else "poorly",
                    previous_aspect="basic scenarios",
                    current_aspect=step.get("sub_aspect", "complex scenarios"),
                    insight="the model's boundaries",
                    topic="model capabilities",
                    tool=step.get("tool", "evaluation tool")
                )
            else:
                template = random.choice(reasoning_templates["summarize"])
                step_copy["enriched_thought"] = template.format(
                    n=i,
                    conclusion="the model has clear strengths and limitations",
                    aspects="all tested dimensions",
                    finding="consistent behavior patterns",
                    key_insight="the model's capability boundaries"
                )
            
            enriched.append(step_copy)
        
        return enriched
    
    def compute_trajectory_metrics(self, trajectory: List[Dict]) -> Dict:
        """Compute quality metrics for a trajectory"""
        metrics = {
            "length": len(trajectory),
            "tool_diversity": len(set(step.get("tool", "") for step in trajectory)),
            "has_summary": any(step.get("decision_type") == "summarize" for step in trajectory),
            "redundancy_score": 0,
            "completeness_score": 0
        }
        
        # Check for redundancy
        aspects = [step.get("sub_aspect", "") for step in trajectory]
        metrics["redundancy_score"] = len(aspects) - len(set(aspects))
        
        # Estimate completeness (heuristic)
        if metrics["length"] < 3:
            metrics["completeness_score"] = 0.3
        elif metrics["length"] > 8:
            metrics["completeness_score"] = 0.7
        else:
            metrics["completeness_score"] = 0.9
        
        return metrics
    
    def create_augmented_dataset(self, output_file: str):
        """Create fully augmented dataset"""
        augmented_data = {
            "version": "2.0",
            "original_examples": 0,
            "augmented_examples": 0,
            "negative_examples": 0,
            "examples": []
        }
        
        # Process each trajectory
        for item in self.base_data.get("trajectories", []):
            trajectory = item.get("trajectory", [])
            user_query = item.get("user_query", "")
            
            # Original example (enriched)
            enriched = self.enrich_with_reasoning(trajectory)
            metrics = self.compute_trajectory_metrics(enriched)
            
            augmented_data["examples"].append({
                "id": f"original_{augmented_data['original_examples']}",
                "user_query": user_query,
                "trajectory": enriched,
                "quality": "good",
                "metrics": metrics,
                "source": "original"
            })
            augmented_data["original_examples"] += 1
            
            # Variations
            variations = self.augment_with_variations(trajectory)
            for var in variations:
                enriched_var = self.enrich_with_reasoning(var)
                metrics = self.compute_trajectory_metrics(enriched_var)
                
                augmented_data["examples"].append({
                    "id": f"augmented_{augmented_data['augmented_examples']}",
                    "user_query": user_query,
                    "trajectory": enriched_var,
                    "quality": "good",
                    "metrics": metrics,
                    "source": "augmented"
                })
                augmented_data["augmented_examples"] += 1
            
            # Negative examples
            negatives = self.add_negative_examples(trajectory)
            for neg in negatives:
                metrics = self.compute_trajectory_metrics(neg["trajectory"])
                
                augmented_data["examples"].append({
                    "id": f"negative_{augmented_data['negative_examples']}",
                    "user_query": user_query,
                    "trajectory": neg["trajectory"],
                    "quality": neg["quality"],
                    "issue": neg["issue"],
                    "metrics": metrics,
                    "source": "negative"
                })
                augmented_data["negative_examples"] += 1
        
        # Save augmented dataset
        with open(output_file, 'w') as f:
            json.dump(augmented_data, f, indent=2)
        
        print(f"Created augmented dataset with:")
        print(f"  Original examples: {augmented_data['original_examples']}")
        print(f"  Augmented examples: {augmented_data['augmented_examples']}")
        print(f"  Negative examples: {augmented_data['negative_examples']}")
        print(f"  Total examples: {len(augmented_data['examples'])}")


# Usage
if __name__ == "__main__":
    augmenter = DatasetAugmenter("collected_trajectories.json")
    augmenter.create_augmented_dataset("augmented_training_data.json")