| | import json |
| | import random |
| | from typing import List, Dict, Any |
| | from base_agent import BaseAgent |
| |
|
| | class TrainingDataGenerator: |
| | """Generate diverse training data for plan agent""" |
| | |
| | def __init__(self): |
| | self.query_templates = self.load_query_templates() |
| | self.tools_t2i = ["Color Binding", "Shape Binding", "Texture Binding", "Non Spatial"] |
| | self.tools_vbench = [ |
| | "Subject Consistency", "Background Consistency", "Motion Smoothness", |
| | "Aesthetic Quality", "Imaging Quality", "Object Class", "Human Action", |
| | "Color", "Spatial Relationship", "Scene" |
| | ] |
| | |
| | def load_query_templates(self) -> Dict[str, List[str]]: |
| | """Load diverse query templates""" |
| | return { |
| | "capability_check": [ |
| | "Can the model generate {object} with {attribute}?", |
| | "How well does the model handle {scenario}?", |
| | "Is the model capable of creating {complex_scene}?" |
| | ], |
| | "comparison": [ |
| | "How does the model perform on {task1} vs {task2}?", |
| | "What are the differences in generating {object1} compared to {object2}?" |
| | ], |
| | "boundary_finding": [ |
| | "What are the limits of the model's ability to {capability}?", |
| | "How complex can {scenario} be before the model fails?", |
| | "What is the maximum number of {elements} the model can handle?" |
| | ], |
| | "quality_assessment": [ |
| | "How consistent is the model in generating {aspect}?", |
| | "What is the quality of {feature} in the generated outputs?", |
| | "How accurate is the model's {binding_type} binding?" |
| | ] |
| | } |
| | |
| | def generate_diverse_queries(self, n: int = 100) -> List[str]: |
| | """Generate diverse evaluation queries""" |
| | queries = [] |
| | |
| | |
| | substitutions = { |
| | "object": ["cats", "cars", "buildings", "people", "landscapes"], |
| | "attribute": ["specific colors", "complex textures", "unusual shapes"], |
| | "scenario": ["multi-object scenes", "dynamic actions", "abstract concepts"], |
| | "complex_scene": ["crowded marketplaces", "underwater scenes", "futuristic cities"], |
| | "task1": ["realistic portraits", "abstract art"], |
| | "task2": ["photorealistic landscapes", "cartoon characters"], |
| | "capability": ["generate multiple objects", "maintain consistency", "follow complex prompts"], |
| | "elements": ["objects", "people", "colors", "textures"], |
| | "aspect": ["human faces", "animal poses", "architectural details"], |
| | "feature": ["motion", "lighting", "composition"], |
| | "binding_type": ["color", "shape", "texture", "spatial"] |
| | } |
| | |
| | for _ in range(n): |
| | template_type = random.choice(list(self.query_templates.keys())) |
| | template = random.choice(self.query_templates[template_type]) |
| | |
| | |
| | query = template |
| | for key, values in substitutions.items(): |
| | if f"{{{key}}}" in query: |
| | query = query.replace(f"{{{key}}}", random.choice(values)) |
| | |
| | queries.append({ |
| | "query": query, |
| | "type": template_type |
| | }) |
| | |
| | return queries |
| | |
| | def generate_exploration_sequence(self, query: str, query_type: str) -> List[Dict]: |
| | """Generate a plausible exploration sequence for a query""" |
| | sequence = [] |
| | |
| | |
| | if query_type == "capability_check": |
| | |
| | complexities = ["simple", "moderate", "complex", "very complex"] |
| | for i, complexity in enumerate(complexities): |
| | sequence.append({ |
| | "step": i + 1, |
| | "sub_aspect": f"Testing with {complexity} scenarios", |
| | "tool": random.choice(self.tools_t2i + self.tools_vbench), |
| | "strategy": "depth-first" |
| | }) |
| | |
| | elif query_type == "comparison": |
| | |
| | aspects = ["first aspect", "second aspect", "combined comparison"] |
| | for i, aspect in enumerate(aspects): |
| | sequence.append({ |
| | "step": i + 1, |
| | "sub_aspect": f"Evaluating {aspect}", |
| | "tool": random.choice(self.tools_t2i + self.tools_vbench), |
| | "strategy": "breadth-first" |
| | }) |
| | |
| | elif query_type == "boundary_finding": |
| | |
| | stress_levels = [10, 50, 90, 99] |
| | for i, level in enumerate(stress_levels): |
| | sequence.append({ |
| | "step": i + 1, |
| | "sub_aspect": f"Testing at {level}th percentile difficulty", |
| | "tool": random.choice(self.tools_t2i + self.tools_vbench), |
| | "strategy": "depth-first" |
| | }) |
| | |
| | else: |
| | |
| | quality_aspects = ["consistency", "accuracy", "diversity", "edge cases"] |
| | for i, aspect in enumerate(quality_aspects): |
| | sequence.append({ |
| | "step": i + 1, |
| | "sub_aspect": f"Assessing {aspect}", |
| | "tool": random.choice(self.tools_t2i + self.tools_vbench), |
| | "strategy": "breadth-first" |
| | }) |
| | |
| | return sequence |
| | |
| | def create_training_example(self, query_data: Dict) -> Dict: |
| | """Create a complete training example""" |
| | query = query_data["query"] |
| | query_type = query_data["type"] |
| | |
| | |
| | exploration = self.generate_exploration_sequence(query, query_type) |
| | |
| | |
| | example = { |
| | "user_query": query, |
| | "query_type": query_type, |
| | "exploration_plan": { |
| | "strategy": "depth-first" if "boundary" in query_type else "breadth-first", |
| | "expected_steps": len(exploration), |
| | "focus_areas": [step["sub_aspect"] for step in exploration] |
| | }, |
| | "exploration_sequence": exploration, |
| | "decision_points": [] |
| | } |
| | |
| | |
| | for i, step in enumerate(exploration): |
| | decision = { |
| | "step": i + 1, |
| | "context": { |
| | "previous_steps": exploration[:i], |
| | "current_observations": f"Simulated results from step {i}" |
| | }, |
| | "decision": "explore" if i < len(exploration) - 1 else "summarize", |
| | "reasoning": f"Need to explore {step['sub_aspect']} to fully answer the query" |
| | } |
| | example["decision_points"].append(decision) |
| | |
| | return example |
| | |
| | def generate_dataset(self, n_examples: int = 1000) -> List[Dict]: |
| | """Generate complete training dataset""" |
| | queries = self.generate_diverse_queries(n_examples) |
| | dataset = [] |
| | |
| | for query_data in queries: |
| | example = self.create_training_example(query_data) |
| | dataset.append(example) |
| | |
| | return dataset |
| |
|
| |
|
| | |
| | if __name__ == "__main__": |
| | generator = TrainingDataGenerator() |
| | |
| | |
| | print("Generating training dataset...") |
| | dataset = generator.generate_dataset(1000) |
| | |
| | |
| | with open("plan_agent_training_data.json", "w") as f: |
| | json.dump({ |
| | "version": "1.0", |
| | "total_examples": len(dataset), |
| | "examples": dataset |
| | }, f, indent=2) |
| | |
| | print(f"Generated {len(dataset)} training examples") |