isa-model 0.4.0__py3-none-any.whl → 0.4.4__py3-none-any.whl

isa_model/training/intelligent/knowledge_base.py

@@ -1,751 +0,0 @@
-"""
-Knowledge Base for Training Intelligence
-
-This module provides a comprehensive knowledge base containing:
-- Model specifications and capabilities
-- Training best practices and benchmarks
-- Performance metrics and cost data
-- Historical training results
-- Resource requirements and recommendations
-
-The knowledge base serves as the brain of the intelligent training system.
-"""
-
-import logging
-import json
-from typing import Dict, List, Optional, Any, Tuple
-from dataclasses import dataclass, field, asdict
-from datetime import datetime
-import os
-from pathlib import Path
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class ModelSpec:
-    """Model specification and metadata."""
-    
-    name: str
-    model_type: str  # "llm", "sd", "cv", "audio"
-    parameters: int  # Number of parameters
-    
-    # Capabilities
-    supported_tasks: List[str]
-    supported_domains: List[str]
-    context_length: int = 2048
-    
-    # Performance characteristics
-    quality_score: float = 0.0  # 0.0 to 1.0
-    speed_score: float = 0.0    # 0.0 to 1.0
-    efficiency_score: float = 0.0  # 0.0 to 1.0
-    
-    # Training characteristics
-    optimal_dataset_size: int = 10000
-    min_dataset_size: int = 100
-    supports_lora: bool = True
-    supports_full_finetuning: bool = True
-    
-    # Resource requirements
-    min_gpu_memory: int = 8  # GB
-    recommended_gpu_memory: int = 16  # GB
-    training_time_factor: float = 1.0  # Relative to baseline
-    
-    # Cost estimates (per hour)
-    estimated_cost_per_hour: float = 0.0
-    
-    # Metadata
-    is_popular: bool = False
-    is_open_source: bool = True
-    release_date: Optional[str] = None
-    description: str = ""
-    
-    # Configuration defaults
-    default_config: Dict[str, Any] = field(default_factory=dict)
-    lora_targets: List[str] = field(default_factory=list)
-    
-    def __post_init__(self):
-        if not self.lora_targets and self.supports_lora:
-            self.lora_targets = ["q_proj", "v_proj", "k_proj", "o_proj"]
-
-
-@dataclass
-class BestPractice:
-    """Training best practice recommendation."""
-    
-    task_type: str
-    domain: str
-    recommendation: str
-    reason: str
-    confidence: float
-    source: str = "expert_knowledge"
-    created_at: datetime = field(default_factory=datetime.now)
-
-
-@dataclass
-class PerformanceBenchmark:
-    """Performance benchmark data."""
-    
-    model_name: str
-    task_type: str
-    dataset_name: str
-    
-    # Performance metrics
-    accuracy: Optional[float] = None
-    bleu_score: Optional[float] = None
-    rouge_score: Optional[float] = None
-    perplexity: Optional[float] = None
-    
-    # Training metrics
-    training_time: Optional[float] = None  # hours
-    training_cost: Optional[float] = None  # USD
-    gpu_type: Optional[str] = None
-    
-    # Configuration used
-    config: Dict[str, Any] = field(default_factory=dict)
-    
-    # Metadata
-    source: str = "benchmark"
-    created_at: datetime = field(default_factory=datetime.now)
-
-
-class KnowledgeBase:
-    """
-    Comprehensive knowledge base for training intelligence.
-    
-    This class manages all the knowledge required for intelligent training decisions:
-    - Model specifications and capabilities
-    - Training best practices and guidelines
-    - Performance benchmarks and historical data
-    - Resource requirements and cost estimates
-    
-    Example:
-        ```python
-        kb = KnowledgeBase()
-        
-        # Get model recommendations
-        models = kb.recommend_models(
-            task_type="chat",
-            domain="medical",
-            dataset_size=5000,
-            quality_target="high"
-        )
-        
-        # Get best practices
-        practices = kb.get_best_practices("chat", "medical")
-        ```
-    """
-    
-    def __init__(self, data_dir: Optional[str] = None):
-        """
-        Initialize knowledge base.
-        
-        Args:
-            data_dir: Directory to store/load knowledge base data
-        """
-        self.data_dir = data_dir or os.path.join(os.getcwd(), "knowledge_base")
-        os.makedirs(self.data_dir, exist_ok=True)
-        
-        # Initialize data structures
-        self.models: Dict[str, ModelSpec] = {}
-        self.best_practices: List[BestPractice] = []
-        self.benchmarks: List[PerformanceBenchmark] = []
-        self.dataset_info: Dict[str, Dict[str, Any]] = {}
-        
-        # Load existing data
-        self._load_knowledge_base()
-        
-        # Initialize with default knowledge if empty
-        if not self.models:
-            self._initialize_default_knowledge()
-        
-        logger.info(f"Knowledge base initialized with {len(self.models)} models")
-    
-    def recommend_models(
-        self,
-        task_type: str,
-        domain: str = "general",
-        dataset_size: int = 10000,
-        quality_target: str = "balanced",
-        constraints: Optional[Dict[str, Any]] = None
-    ) -> List[Dict[str, Any]]:
-        """
-        Recommend models for a specific task and requirements.
-        
-        Args:
-            task_type: Type of task (chat, classification, etc.)
-            domain: Domain/industry
-            dataset_size: Size of training dataset
-            quality_target: Quality target (fast, balanced, high)
-            constraints: Additional constraints (budget, time, etc.)
-            
-        Returns:
-            List of recommended models with scores
-        """
-        constraints = constraints or {}
-        
-        # Filter models by task compatibility
-        compatible_models = []
-        for model_name, model_spec in self.models.items():
-            if self._is_model_compatible(model_spec, task_type, domain, dataset_size):
-                score = self._score_model(
-                    model_spec, task_type, domain, dataset_size, quality_target, constraints
-                )
-                
-                model_dict = asdict(model_spec)
-                model_dict["suitability_score"] = score
-                model_dict["task_suitability"] = score / 100.0  # Normalize
-                
-                # Ensure trainer_type is set based on model_type
-                if "trainer_type" not in model_dict or not model_dict["trainer_type"]:
-                    if model_spec.model_type == "llm":
-                        model_dict["trainer_type"] = "llm"
-                    elif model_spec.model_type == "sd":
-                        model_dict["trainer_type"] = "sd"
-                    elif model_spec.model_type == "ml":
-                        model_dict["trainer_type"] = "ml"
-                    else:
-                        model_dict["trainer_type"] = "llm"  # Default
-                
-                compatible_models.append(model_dict)
-        
-        # Sort by suitability score
-        compatible_models.sort(key=lambda x: x["suitability_score"], reverse=True)
-        
-        return compatible_models[:10]  # Return top 10
-    
-    def _is_model_compatible(
-        self, 
-        model_spec: ModelSpec, 
-        task_type: str, 
-        domain: str, 
-        dataset_size: int
-    ) -> bool:
-        """Check if model is compatible with requirements."""
-        # Check task compatibility
-        if task_type not in model_spec.supported_tasks and "general" not in model_spec.supported_tasks:
-            return False
-        
-        # Check domain compatibility
-        if domain not in model_spec.supported_domains and "general" not in model_spec.supported_domains:
-            return False
-        
-        # Check minimum dataset size
-        if dataset_size < model_spec.min_dataset_size:
-            return False
-        
-        return True
-    
-    def _score_model(
-        self,
-        model_spec: ModelSpec,
-        task_type: str,
-        domain: str,
-        dataset_size: int,
-        quality_target: str,
-        constraints: Dict[str, Any]
-    ) -> float:
-        """Score a model's suitability for the task."""
-        score = 0.0
-        
-        # Base compatibility score
-        if task_type in model_spec.supported_tasks:
-            score += 30
-        elif "general" in model_spec.supported_tasks:
-            score += 15
-        
-        if domain in model_spec.supported_domains:
-            score += 20
-        elif "general" in model_spec.supported_domains:
-            score += 10
-        
-        # Quality vs speed preference
-        if quality_target == "high":
-            score += model_spec.quality_score * 25
-        elif quality_target == "fast":
-            score += model_spec.speed_score * 25
-        else:  # balanced
-            score += (model_spec.quality_score + model_spec.speed_score) * 12.5
-        
-        # Efficiency bonus
-        score += model_spec.efficiency_score * 10
-        
-        # Popularity bonus
-        if model_spec.is_popular:
-            score += 5
-        
-        # Dataset size compatibility
-        if dataset_size >= model_spec.optimal_dataset_size:
-            score += 10
-        else:
-            # Penalty for suboptimal dataset size
-            ratio = dataset_size / model_spec.optimal_dataset_size
-            score += ratio * 10
-        
-        # Budget constraints
-        if constraints.get("budget"):
-            budget = constraints["budget"]
-            estimated_cost = model_spec.estimated_cost_per_hour * 10  # Assume 10 hours
-            if estimated_cost <= budget:
-                score += 5
-            else:
-                score -= 10
-        
-        # Time constraints
-        if constraints.get("time"):
-            time_limit = constraints["time"]
-            estimated_time = 10 * model_spec.training_time_factor  # Assume 10 base hours
-            if estimated_time <= time_limit:
-                score += 5
-            else:
-                score -= 10
-        
-        # Model preferences
-        if constraints.get("preferences"):
-            preferences = constraints["preferences"]
-            if preferences:
-                for pref in preferences:
-                    if pref.lower() in model_spec.name.lower():
-                        score += 15
-        
-        return max(0.0, score)
-    
-    def get_best_practices(self, task_type: str, domain: str = "general") -> List[BestPractice]:
-        """Get best practices for a specific task and domain."""
-        practices = []
-        
-        for practice in self.best_practices:
-            if (practice.task_type == task_type or practice.task_type == "general") and \
-               (practice.domain == domain or practice.domain == "general"):
-                practices.append(practice)
-        
-        # Sort by confidence
-        practices.sort(key=lambda x: x.confidence, reverse=True)
-        
-        return practices
-    
-    def get_performance_benchmarks(
-        self, 
-        model_name: Optional[str] = None,
-        task_type: Optional[str] = None
-    ) -> List[PerformanceBenchmark]:
-        """Get performance benchmarks."""
-        benchmarks = []
-        
-        for benchmark in self.benchmarks:
-            if model_name and benchmark.model_name != model_name:
-                continue
-            if task_type and benchmark.task_type != task_type:
-                continue
-            benchmarks.append(benchmark)
-        
-        return benchmarks
-    
-    def get_dataset_info(self, dataset_name: str) -> Optional[Dict[str, Any]]:
-        """Get information about a dataset."""
-        return self.dataset_info.get(dataset_name)
-    
-    def add_model(self, model_spec: ModelSpec) -> None:
-        """Add a new model to the knowledge base."""
-        self.models[model_spec.name] = model_spec
-        self._save_models()
-        logger.info(f"Added model to knowledge base: {model_spec.name}")
-    
-    def add_best_practice(self, practice: BestPractice) -> None:
-        """Add a best practice to the knowledge base."""
-        self.best_practices.append(practice)
-        self._save_best_practices()
-        logger.info(f"Added best practice: {practice.task_type}/{practice.domain}")
-    
-    def add_benchmark(self, benchmark: PerformanceBenchmark) -> None:
-        """Add a performance benchmark."""
-        self.benchmarks.append(benchmark)
-        self._save_benchmarks()
-        logger.info(f"Added benchmark: {benchmark.model_name}/{benchmark.task_type}")
-    
-    def update_from_training_result(self, training_result: Dict[str, Any]) -> None:
-        """Update knowledge base from training results."""
-        # Extract information from training result
-        model_name = training_result.get("model_name")
-        task_type = training_result.get("task_type")
-        
-        if model_name and task_type:
-            # Create benchmark from result
-            benchmark = PerformanceBenchmark(
-                model_name=model_name,
-                task_type=task_type,
-                dataset_name=training_result.get("dataset_name", "unknown"),
-                training_time=training_result.get("training_time"),
-                training_cost=training_result.get("training_cost"),
-                gpu_type=training_result.get("gpu_type"),
-                config=training_result.get("config", {}),
-                source="training_result"
-            )
-            
-            self.add_benchmark(benchmark)
-    
-    def _initialize_default_knowledge(self) -> None:
-        """Initialize knowledge base with default models and best practices."""
-        # Add popular LLM models
-        self._add_llm_models()
-        
-        # Add Stable Diffusion models
-        self._add_sd_models()
-        
-        # Add ML models
-        self._add_ml_models()
-        
-        # Add best practices
-        self._add_default_best_practices()
-        
-        # Add dataset information
-        self._add_default_dataset_info()
-        
-        # Save to disk
-        self._save_knowledge_base()
-        
-        logger.info("Initialized knowledge base with default data")
-    
-    def _add_llm_models(self) -> None:
-        """Add LLM models to knowledge base."""
-        llm_models = [
-            ModelSpec(
-                name="google/gemma-2-2b-it",
-                model_type="llm",
-                parameters=2_000_000_000,
-                supported_tasks=["chat", "classification", "generation", "summarization"],
-                supported_domains=["general", "technical", "education"],
-                context_length=2048,
-                quality_score=0.7,
-                speed_score=0.9,
-                efficiency_score=0.8,
-                optimal_dataset_size=5000,
-                min_dataset_size=100,
-                min_gpu_memory=8,
-                recommended_gpu_memory=12,
-                training_time_factor=0.6,
-                estimated_cost_per_hour=0.5,
-                is_popular=True,
-                is_open_source=True,
-                description="Lightweight Gemma model optimized for efficiency",
-                default_config={"learning_rate": 2e-5, "batch_size": 4, "lora_rank": 8}
-            ),
-            ModelSpec(
-                name="google/gemma-2-4b-it",
-                model_type="llm",
-                parameters=4_000_000_000,
-                supported_tasks=["chat", "classification", "generation", "summarization", "reasoning"],
-                supported_domains=["general", "technical", "education", "medical"],
-                context_length=2048,
-                quality_score=0.8,
-                speed_score=0.7,
-                efficiency_score=0.8,
-                optimal_dataset_size=10000,
-                min_dataset_size=200,
-                min_gpu_memory=12,
-                recommended_gpu_memory=16,
-                training_time_factor=0.8,
-                estimated_cost_per_hour=0.8,
-                is_popular=True,
-                is_open_source=True,
-                description="Balanced Gemma model with good performance and efficiency",
-                default_config={"learning_rate": 2e-5, "batch_size": 2, "lora_rank": 16}
-            ),
-            ModelSpec(
-                name="google/gemma-2-7b-it",
-                model_type="llm",
-                parameters=7_000_000_000,
-                supported_tasks=["chat", "classification", "generation", "summarization", "reasoning", "code"],
-                supported_domains=["general", "technical", "education", "medical", "legal"],
-                context_length=4096,
-                quality_score=0.9,
-                speed_score=0.6,
-                efficiency_score=0.7,
-                optimal_dataset_size=20000,
-                min_dataset_size=500,
-                min_gpu_memory=16,
-                recommended_gpu_memory=24,
-                training_time_factor=1.0,
-                estimated_cost_per_hour=1.2,
-                is_popular=True,
-                is_open_source=True,
-                description="High-quality Gemma model for demanding tasks",
-                default_config={"learning_rate": 1e-5, "batch_size": 1, "lora_rank": 32}
-            ),
-            ModelSpec(
-                name="microsoft/DialoGPT-medium",
-                model_type="llm",
-                parameters=345_000_000,
-                supported_tasks=["chat", "generation"],
-                supported_domains=["general", "customer_service"],
-                context_length=1024,
-                quality_score=0.6,
-                speed_score=0.95,
-                efficiency_score=0.9,
-                optimal_dataset_size=2000,
-                min_dataset_size=50,
-                min_gpu_memory=4,
-                recommended_gpu_memory=8,
-                training_time_factor=0.3,
-                estimated_cost_per_hour=0.3,
-                is_popular=True,
-                is_open_source=True,
-                description="Fast conversational model for chatbots",
-                default_config={"learning_rate": 5e-5, "batch_size": 8, "lora_rank": 4}
-            ),
-            ModelSpec(
-                name="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
-                model_type="llm",
-                parameters=1_100_000_000,
-                supported_tasks=["chat", "generation"],
-                supported_domains=["general", "education"],
-                context_length=2048,
-                quality_score=0.5,
-                speed_score=0.95,
-                efficiency_score=0.9,
-                optimal_dataset_size=1000,
-                min_dataset_size=50,
-                min_gpu_memory=4,
-                recommended_gpu_memory=6,
-                training_time_factor=0.4,
-                estimated_cost_per_hour=0.2,
-                is_popular=True,
-                is_open_source=True,
-                description="Extremely lightweight model for resource-constrained environments",
-                default_config={"learning_rate": 3e-5, "batch_size": 8, "lora_rank": 8}
-            )
-        ]
-        
-        for model in llm_models:
-            self.models[model.name] = model
-    
-    def _add_sd_models(self) -> None:
-        """Add Stable Diffusion models."""
-        sd_models = [
-            ModelSpec(
-                name="runwayml/stable-diffusion-v1-5",
-                model_type="sd",
-                parameters=860_000_000,
-                supported_tasks=["image_generation", "image_editing"],
-                supported_domains=["art", "design", "general"],
-                context_length=77,
-                quality_score=0.8,
-                speed_score=0.6,
-                efficiency_score=0.7,
-                optimal_dataset_size=1000,
-                min_dataset_size=20,
-                min_gpu_memory=8,
-                recommended_gpu_memory=12,
-                training_time_factor=1.2,
-                estimated_cost_per_hour=1.0,
-                is_popular=True,
-                is_open_source=True,
-                description="Popular Stable Diffusion model for image generation",
-                default_config={"learning_rate": 1e-6, "batch_size": 1}
-            )
-        ]
-        
-        for model in sd_models:
-            self.models[model.name] = model
-    
-    def _add_ml_models(self) -> None:
-        """Add traditional ML models."""
-        ml_models = [
-            ModelSpec(
-                name="xgboost_classifier",
-                model_type="ml",
-                parameters=1000,  # Approximate
-                supported_tasks=["classification", "regression"],
-                supported_domains=["general", "financial", "medical"],
-                context_length=0,
-                quality_score=0.8,
-                speed_score=0.9,
-                efficiency_score=0.95,
-                optimal_dataset_size=10000,
-                min_dataset_size=100,
-                min_gpu_memory=2,
-                recommended_gpu_memory=4,
-                training_time_factor=0.1,
-                estimated_cost_per_hour=0.1,
-                is_popular=True,
-                is_open_source=True,
-                description="Gradient boosting classifier for tabular data",
-                default_config={"n_estimators": 100, "learning_rate": 0.1}
-            ),
-            ModelSpec(
-                name="random_forest_classifier",
-                model_type="ml",
-                parameters=1000,
-                supported_tasks=["classification", "regression"],
-                supported_domains=["general", "financial", "medical"],
-                context_length=0,
-                quality_score=0.7,
-                speed_score=0.8,
-                efficiency_score=0.9,
-                optimal_dataset_size=5000,
-                min_dataset_size=100,
-                min_gpu_memory=2,
-                recommended_gpu_memory=4,
-                training_time_factor=0.2,
-                estimated_cost_per_hour=0.05,
-                is_popular=True,
-                is_open_source=True,
-                description="Random forest classifier for tabular data",
-                default_config={"n_estimators": 100, "max_depth": 10}
-            )
-        ]
-        
-        for model in ml_models:
-            self.models[model.name] = model
-    
-    def _add_default_best_practices(self) -> None:
-        """Add default best practices."""
-        practices = [
-            BestPractice(
-                task_type="chat",
-                domain="general",
-                recommendation="Use LoRA with rank 8-16 for chat models",
-                reason="LoRA provides good performance while reducing training time and memory usage",
-                confidence=0.9
-            ),
-            BestPractice(
-                task_type="chat",
-                domain="medical",
-                recommendation="Use higher learning rates (3e-5) for medical domain adaptation",
-                reason="Medical terminology requires stronger adaptation from general models",
-                confidence=0.8
-            ),
-            BestPractice(
-                task_type="classification",
-                domain="general",
-                recommendation="Use batch size 8-16 for classification tasks",
-                reason="Larger batch sizes improve stability for classification training",
-                confidence=0.85
-            ),
-            BestPractice(
-                task_type="generation",
-                domain="general",
-                recommendation="Use gradient accumulation for large models",
-                reason="Maintains effective batch size while reducing memory usage",
-                confidence=0.9
-            )
-        ]
-        
-        self.best_practices.extend(practices)
-    
-    def _add_default_dataset_info(self) -> None:
-        """Add default dataset information."""
-        self.dataset_info = {
-            "tatsu-lab/alpaca": {
-                "size": 52000,
-                "format": "alpaca",
-                "language": "english",
-                "domain": "general",
-                "quality": "high"
-            },
-            "Open-Orca/OpenOrca": {
-                "size": 4200000,
-                "format": "sharegpt",
-                "language": "english",
-                "domain": "general",
-                "quality": "high"
-            },
-            "microsoft/DialoGPT-medium": {
-                "size": 147000000,
-                "format": "conversational",
-                "language": "english",
-                "domain": "general",
-                "quality": "medium"
-            }
-        }
-    
-    def _load_knowledge_base(self) -> None:
-        """Load knowledge base from disk."""
-        try:
-            self._load_models()
-            self._load_best_practices()
-            self._load_benchmarks()
-            self._load_dataset_info()
-        except Exception as e:
-            logger.warning(f"Failed to load knowledge base: {e}")
-    
-    def _save_knowledge_base(self) -> None:
-        """Save knowledge base to disk."""
-        try:
-            self._save_models()
-            self._save_best_practices()
-            self._save_benchmarks()
-            self._save_dataset_info()
-        except Exception as e:
-            logger.error(f"Failed to save knowledge base: {e}")
-    
-    def _load_models(self) -> None:
-        """Load models from disk."""
-        models_file = os.path.join(self.data_dir, "models.json")
-        if os.path.exists(models_file):
-            with open(models_file, 'r') as f:
-                data = json.load(f)
-                for name, model_data in data.items():
-                    self.models[name] = ModelSpec(**model_data)
-    
-    def _save_models(self) -> None:
-        """Save models to disk."""
-        models_file = os.path.join(self.data_dir, "models.json")
-        with open(models_file, 'w') as f:
-            data = {name: asdict(model) for name, model in self.models.items()}
-            json.dump(data, f, indent=2, default=str)
-    
-    def _load_best_practices(self) -> None:
-        """Load best practices from disk."""
-        practices_file = os.path.join(self.data_dir, "best_practices.json")
-        if os.path.exists(practices_file):
-            with open(practices_file, 'r') as f:
-                data = json.load(f)
-                self.best_practices = [BestPractice(**item) for item in data]
-    
-    def _save_best_practices(self) -> None:
-        """Save best practices to disk."""
-        practices_file = os.path.join(self.data_dir, "best_practices.json")
-        with open(practices_file, 'w') as f:
-            data = [asdict(practice) for practice in self.best_practices]
-            json.dump(data, f, indent=2, default=str)
-    
-    def _load_benchmarks(self) -> None:
-        """Load benchmarks from disk."""
-        benchmarks_file = os.path.join(self.data_dir, "benchmarks.json")
-        if os.path.exists(benchmarks_file):
-            with open(benchmarks_file, 'r') as f:
-                data = json.load(f)
-                self.benchmarks = [PerformanceBenchmark(**item) for item in data]
-    
-    def _save_benchmarks(self) -> None:
-        """Save benchmarks to disk."""
-        benchmarks_file = os.path.join(self.data_dir, "benchmarks.json")
-        with open(benchmarks_file, 'w') as f:
-            data = [asdict(benchmark) for benchmark in self.benchmarks]
-            json.dump(data, f, indent=2, default=str)
-    
-    def _load_dataset_info(self) -> None:
-        """Load dataset info from disk."""
-        dataset_file = os.path.join(self.data_dir, "datasets.json")
-        if os.path.exists(dataset_file):
-            with open(dataset_file, 'r') as f:
-                self.dataset_info = json.load(f)
-    
-    def _save_dataset_info(self) -> None:
-        """Save dataset info to disk."""
-        dataset_file = os.path.join(self.data_dir, "datasets.json")
-        with open(dataset_file, 'w') as f:
-            json.dump(self.dataset_info, f, indent=2)
-    
-    def get_statistics(self) -> Dict[str, Any]:
-        """Get knowledge base statistics."""
-        return {
-            "total_models": len(self.models),
-            "llm_models": len([m for m in self.models.values() if m.model_type == "llm"]),
-            "sd_models": len([m for m in self.models.values() if m.model_type == "sd"]),
-            "ml_models": len([m for m in self.models.values() if m.model_type == "ml"]),
-            "best_practices": len(self.best_practices),
-            "benchmarks": len(self.benchmarks),
-            "datasets": len(self.dataset_info)
-        }