process_hf_dataset.py

# process_hf_dataset.py
from datasets import load_dataset
import re
from parser import parse_python_code, create_vector
from database import init_chromadb, store_program, DB_NAME, HF_DATASET_NAME
import chromadb
import os
from dotenv import load_dotenv

# Load environment variables
load_dotenv()

def rename_variables(code, variable_prefixes=None):
    """Rename variables in Python code to align with vector categories (input_variable, assigned_variable, returned_variable)."""
    if variable_prefixes is None:
        variable_prefixes = {
            'input': 'input_variable',
            'assigned': 'assigned_variable',
            'returned': 'returned_variable'
        }
    
    # Simple variable name detection and renaming
    pattern = r'\b[a-zA-Z_]\w*\b'  # Match variable names (simple heuristic)
    variables = set()
    code_lines = code.split('\n')
    
    # Find all variable names (simplified approach, could improve with AST)
    for line in code_lines:
        matches = re.findall(pattern, line)
        for match in matches:
            if match not in ['def', 'if', 'else', 'for', 'while', 'return', 'import', 'print', 'eval', 'str', 'int']:  # Exclude keywords
                variables.add(match)
    
    # Sort variables by first appearance (simplified, could improve with AST)
    sorted_vars = sorted(list(variables))
    var_map = {}
    var_count = {'input_variable': 1, 'assigned_variable': 1, 'returned_variable': 1}
    
    # Assign variables based on context (simplified heuristic)
    for var in sorted_vars:
        # Determine variable role based on context
        is_input = any(var in line and 'def' in line for line in code_lines)  # Check if in function definition (input parameter)
        is_returned = any('return' in line and var in line for line in code_lines)  # Check if used in return statement
        is_assigned = any('=' in line and var in line.split('=')[0].strip() for line in code_lines)  # Check if assigned
        
        if is_input:
            role = 'input_variable'
        elif is_returned:
            role = 'returned_variable'
        elif is_assigned:
            role = 'assigned_variable'
        else:
            role = 'assigned_variable'  # Default to assigned if unclear
        
        new_name = f"{role}{var_count[role]}"
        var_map[var] = new_name
        var_count[role] += 1
    
    # Replace variables in code
    new_code = code
    for old_var, new_var in var_map.items():
        new_code = re.sub(r'\b' + old_var + r'\b', new_var, new_code)
    
    return new_code, var_map

def generate_description_tokens(sequence, vectors, var_map=None):
    """Generate semantic description tokens for a program, including variable roles."""
    tokens = []
    category_descriptions = {
        'import': 'imports module',
        'function': 'defines function',
        'assigned_variable': 'assigns variable',
        'input_variable': 'input parameter',
        'returned_variable': 'returns value',
        'if': 'conditional statement',
        'return': 'returns result',
        'try': 'try block',
        'except': 'exception handler',
        'expression': 'expression statement',
        'spacer': 'empty line or comment'
    }
    
    for cat, vec in zip(sequence, vectors):
        if cat in category_descriptions:
            tokens.append(f"{category_descriptions[cat]}:{cat}")
            # Add vector-derived features (e.g., level, span) as tokens
            tokens.append(f"level:{vec[1]}")
            tokens.append(f"span:{vec[3]:.2f}")
    
    # Add variable role tokens if var_map exists
    if var_map:
        for old_var, new_var in var_map.items():
            role = new_var.split('variable')[0] + 'variable'  # Extract role (e.g., 'input_variable')
            tokens.append(f"variable:{old_var}={new_var}:{role}")
    
    return tokens

def generate_semantic_vector_og(description, total_lines=100):
    """Generate a 6D semantic vector for a textual description, matching our vector format."""
    # Use a simplified heuristic to map description to our 6D vector format
    category_map = {
        'import': 1, 'function': 2, 'assign': 17, 'input': 18, 'return': 19, 'if': 5, 'try': 8, 'except': 14
    }
    
    # Parse description for key terms
    tokens = description.lower().split()
    vector = [0] * 6  # Initialize 6D vector
    
    # Map description tokens to categories and assign basic vector values
    for token in tokens:
        for cat, cat_id in category_map.items():
            if cat in token:
                vector[0] = cat_id  # category_id
                vector[1] = 1  # level (assume top-level for simplicity)
                vector[2] = 0.5  # center_pos (midpoint of code)
                vector[3] = 0.1  # span (small for simplicity)
                vector[4] = 1  # parent_depth (shallow)
                vector[5] = cat_id / len(category_map)  # parent_weight (normalized)
                break
    
    return vector


"""Generate a 6D semantic vector for a textual description using CodeBERT, projecting to 6D."""
# Load CodeBERT model and tokenizer
model_name = "microsoft/codebert-base"
tokenizer = AutoTokenizer.from_pretrained(model_name)
device = torch.device("cuda" if use_gpu and torch.cuda.is_available() else "cpu")
model = AutoModel.from_pretrained(model_name).to(device)
def generate_semantic_vector(description, total_lines=100, use_gpu=False):

    
    # Tokenize and encode the description
    inputs = tokenizer(description, return_tensors="pt", padding=True, truncation=True, max_length=512)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    # Generate embeddings
    with torch.no_grad():
        outputs = model(**inputs)
        # Use mean pooling of the last hidden states
        vector = outputs.last_hidden_state.mean(dim=1).squeeze().cpu().numpy().tolist()
    
    # Truncate or project to 6D (simplified projection: take first 6 dimensions)
    if len(vector) < 6:
        vector.extend([0] * (6 - len(vector)))
    elif len(vector) > 6:
        vector = vector[:6]  # Truncate to 6D
    
    return vector
def process_hf_dataset():
    """Process the Hugging Face dataset and store programs in ChromaDB, aligning with vector categories and including instruction in vectors."""
    # Load the dataset
    dataset = load_dataset("iamtarun/python_code_instructions_18k_alpaca", split="train")
    
    # Initialize ChromaDB client
    client = init_chromadb()
    
    # Process each entry
    for entry in dataset:
        instruction = entry['instruction']
        output = entry['output']
        
        # Rename variables to align with vector categories
        processed_code, var_map = rename_variables(output)
        
        # Parse the code to get parts and sequence, generating our 6D vectors
        parts, sequence = parse_python_code(processed_code)
        vectors = [part['vector'] for part in parts]  # Use parser's 6D vectors
        
        # Generate description tokens including variable roles
        description_tokens = f"task:{instruction.replace(' ', '_')}"
        description_tokens_list = generate_description_tokens(sequence, vectors, var_map)
        description_tokens += " " + " ".join(description_tokens_list)
        
        # Generate a 6D semantic vector for the instruction, incorporating it into the program vector
        semantic_vector = generate_semantic_vector(instruction)
        
        # Combine program vectors with instruction vector (average or concatenate, but ensure 6D)
        combined_vector = semantic_vector  # Use semantic vector as primary for semantic search
        
        # Store in ChromaDB with description and combined vector
        store_program(client, processed_code, sequence, [combined_vector], DB_NAME)
        
        # Update metadata with instruction and variable roles as description
        collection = client.get_collection(DB_NAME)
        program_id = str(hash(processed_code))
        collection.update(
            ids=[program_id],
            metadatas=[{"sequence": ",".join(sequence), "description_tokens": description_tokens}],
            embeddings=[combined_vector]  # Ensure 6D embedding
        )
    
    # Save to Hugging Face Dataset
    save_chromadb_to_hf()

def save_chromadb_to_hf(dataset_name=HF_DATASET_NAME, token=os.getenv("HF_KEY")):
    """Save ChromaDB data to Hugging Face Dataset."""
    client = init_chromadb()
    collection = client.get_collection(DB_NAME)
    
    # Fetch all data from ChromaDB
    results = collection.get(include=["documents", "metadatas", "embeddings"])
    data = {
        "code": results["documents"],
        "sequence": [meta["sequence"] for meta in results["metadatas"]],
        "vectors": results["embeddings"],  # ChromaDB already flattens embeddings
        "description_tokens": [meta.get('description_tokens', '') for meta in results["metadatas"]]
    }
    
    # Create a Hugging Face Dataset
    dataset = Dataset.from_dict(data)
    
    # Push to Hugging Face Hub
    dataset.push_to_hub(dataset_name, token=token)
    print(f"Dataset pushed to Hugging Face Hub as {dataset_name}")

if __name__ == "__main__":
    process_hf_dataset()