# 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 from transformers import AutoTokenizer, AutoModel import torch # 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(description, total_lines=100, use_gpu=False): """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) # 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) program_vectors = [part['vector'] for part in parts] # Use parser's 6D vectors for program structure # Generate description tokens including variable roles description_tokens = f"task:{instruction.replace(' ', '_')}" description_tokens_list = generate_description_tokens(sequence, program_vectors, var_map) description_tokens += " " + " ".join(description_tokens_list) # Generate a 6D semantic vector for the instruction semantic_vector = generate_semantic_vector(instruction) # Combine program vectors with semantic vector (use semantic vector for semantic search, store program vectors separately) # Store both semantic and program vectors, but ensure ChromaDB uses 6D combined_vector = semantic_vector # Use semantic vector for ChromaDB embedding (6D) # 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, and store program vectors 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, "program_vectors": str(program_vectors)}], embeddings=[combined_vector] # Ensure 6D embedding for semantic search ) # 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"], # Semantic 6D vectors "description_tokens": [meta.get('description_tokens', '') for meta in results["metadatas"]], "program_vectors": [eval(meta.get('program_vectors', '[]')) for meta in results["metadatas"]] # Store structural vectors } # 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()