isa-model 0.2.0__py3-none-any.whl → 0.2.8__py3-none-any.whl

isa_model/training/core/dataset.py

@@ -0,0 +1,222 @@
+"""
+Dataset Management
+
+Handles loading and preprocessing of datasets for training.
+"""
+
+import json
+import logging
+from typing import Optional, Tuple, Dict, Any, List, Union
+from pathlib import Path
+from datasets import Dataset, load_dataset
+
+logger = logging.getLogger(__name__)
+
+
+class DatasetManager:
+    """Manages dataset loading and preprocessing."""
+    
+    def __init__(self, tokenizer, max_length: int = 1024):
+        """
+        Initialize dataset manager.
+        
+        Args:
+            tokenizer: HuggingFace tokenizer
+            max_length: Maximum sequence length
+        """
+        self.tokenizer = tokenizer
+        self.max_length = max_length
+        
+    def prepare_dataset(
+        self,
+        dataset_path: str,
+        dataset_format: str = "alpaca",
+        validation_split: float = 0.1
+    ) -> Tuple[Dataset, Optional[Dataset]]:
+        """
+        Prepare training and validation datasets.
+        
+        Args:
+            dataset_path: Path to dataset file or HuggingFace dataset name
+            dataset_format: Format of the dataset (alpaca, sharegpt, custom)
+            validation_split: Fraction of data to use for validation
+            
+        Returns:
+            Tuple of (train_dataset, eval_dataset)
+        """
+        logger.info(f"Preparing dataset: {dataset_path}")
+        
+        # Load raw dataset
+        raw_dataset = self._load_raw_dataset(dataset_path)
+        
+        # Convert to training format
+        if dataset_format == "alpaca":
+            processed_dataset = self._process_alpaca_format(raw_dataset)
+        elif dataset_format == "sharegpt":
+            processed_dataset = self._process_sharegpt_format(raw_dataset)
+        else:
+            processed_dataset = self._process_custom_format(raw_dataset)
+        
+        # Tokenize dataset
+        tokenized_dataset = processed_dataset.map(
+            self._tokenize_function,
+            batched=True,
+            remove_columns=processed_dataset.column_names
+        )
+        
+        # Split into train/eval
+        if validation_split > 0:
+            split_dataset = tokenized_dataset.train_test_split(
+                test_size=validation_split,
+                seed=42
+            )
+            train_dataset = split_dataset["train"]
+            eval_dataset = split_dataset["test"]
+        else:
+            train_dataset = tokenized_dataset
+            eval_dataset = None
+        
+        logger.info(f"Dataset prepared: {len(train_dataset)} training samples")
+        if eval_dataset:
+            logger.info(f"Validation samples: {len(eval_dataset)}")
+        
+        return train_dataset, eval_dataset
+    
+    def _load_raw_dataset(self, dataset_path: str) -> Dataset:
+        """Load raw dataset from file or HuggingFace."""
+        try:
+            # Check if it's a local file
+            if Path(dataset_path).exists():
+                logger.info(f"Loading local dataset: {dataset_path}")
+                
+                if dataset_path.endswith('.json'):
+                    with open(dataset_path, 'r') as f:
+                        data = json.load(f)
+                    return Dataset.from_list(data)
+                elif dataset_path.endswith('.jsonl'):
+                    data = []
+                    with open(dataset_path, 'r') as f:
+                        for line in f:
+                            data.append(json.loads(line))
+                    return Dataset.from_list(data)
+                else:
+                    raise ValueError(f"Unsupported file format: {dataset_path}")
+            
+            else:
+                # Try loading from HuggingFace Hub
+                logger.info(f"Loading HuggingFace dataset: {dataset_path}")
+                dataset = load_dataset(dataset_path, split="train")
+                return dataset
+                
+        except Exception as e:
+            logger.error(f"Failed to load dataset: {e}")
+            raise
+    
+    def _process_alpaca_format(self, dataset: Dataset) -> Dataset:
+        """Process Alpaca format dataset."""
+        def format_alpaca(example):
+            instruction = example.get("instruction", "")
+            input_text = example.get("input", "")
+            output = example.get("output", "")
+            
+            # Format prompt
+            if input_text:
+                prompt = f"### Instruction:\n{instruction}\n\n### Input:\n{input_text}\n\n### Response:\n"
+            else:
+                prompt = f"### Instruction:\n{instruction}\n\n### Response:\n"
+            
+            # Combine prompt and response
+            text = prompt + output
+            
+            return {"text": text}
+        
+        return dataset.map(format_alpaca)
+    
+    def _process_sharegpt_format(self, dataset: Dataset) -> Dataset:
+        """Process ShareGPT format dataset."""
+        def format_sharegpt(example):
+            conversations = example.get("conversations", [])
+            
+            text = ""
+            for conv in conversations:
+                role = conv.get("from", "")
+                content = conv.get("value", "")
+                
+                if role == "human":
+                    text += f"### Human:\n{content}\n\n"
+                elif role == "gpt":
+                    text += f"### Assistant:\n{content}\n\n"
+            
+            return {"text": text.strip()}
+        
+        return dataset.map(format_sharegpt)
+    
+    def _process_custom_format(self, dataset: Dataset) -> Dataset:
+        """Process custom format dataset."""
+        # Assume the dataset already has a 'text' column
+        if "text" not in dataset.column_names:
+            raise ValueError("Custom format dataset must have a 'text' column")
+        
+        return dataset
+    
+    def _tokenize_function(self, examples):
+        """Tokenize examples for training."""
+        # Tokenize inputs
+        tokenized = self.tokenizer(
+            examples["text"],
+            truncation=True,
+            padding=False,
+            max_length=self.max_length,
+            return_tensors=None,
+        )
+        
+        # For language modeling, labels are the same as input_ids
+        tokenized["labels"] = tokenized["input_ids"].copy()
+        
+        return tokenized
+    
+    @staticmethod
+    def convert_hf_dataset_to_alpaca(
+        dataset_name: str,
+        output_path: str,
+        instruction_column: str = "instruction",
+        input_column: str = "input",
+        output_column: str = "output"
+    ) -> str:
+        """
+        Convert a HuggingFace dataset to Alpaca format.
+        
+        Args:
+            dataset_name: Name of the HuggingFace dataset
+            output_path: Path to save the converted dataset
+            instruction_column: Column name for instructions
+            input_column: Column name for inputs
+            output_column: Column name for outputs
+            
+        Returns:
+            Path to the saved dataset
+        """
+        logger.info(f"Converting {dataset_name} to Alpaca format")
+        
+        # Load dataset
+        dataset = load_dataset(dataset_name, split="train")
+        
+        # Convert to Alpaca format
+        alpaca_data = []
+        for example in dataset:
+            alpaca_example = {
+                "instruction": example.get(instruction_column, ""),
+                "input": example.get(input_column, ""),
+                "output": example.get(output_column, "")
+            }
+            alpaca_data.append(alpaca_example)
+        
+        # Save to file
+        output_path = Path(output_path)
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        
+        with open(output_path, 'w') as f:
+            json.dump(alpaca_data, f, indent=2)
+        
+        logger.info(f"Dataset converted and saved to: {output_path}")
+        return str(output_path)