trl copy.py

#!/usr/bin/env python3
"""
NZ Legislation Loophole Analysis Dataset Creation Tool

This script processes New Zealand legislation text to create a finetuning dataset for AI models
that can identify potential loopholes, ambiguities, and unintended consequences in legal text.

The script:
1. Loads and cleans NZ legislation text, preserving legal structure and terminology
2. Chunks the text into manageable sections with overlap for context
3. Uses an LLM to analyze each chunk for legal issues
4. Generates a structured dataset for training AI models on legal loophole detection

Usage:
    python trl.py

Requirements:
    - llama-cpp-python with GGUF model support
    - psutil for memory monitoring
    - Input file: nz-legislation.txt containing NZ legislation in JSON lines format

Output:
    - JSON dataset saved to nz_legislation_dataset/nz_legislation_loophole_dataset.json
"""

import os
import json
import time
import psutil
from typing import List, Dict, Any
import numpy as np
from llama_cpp import Llama
import re

# Placeholder classes and functions for missing dependencies
class ProgressManager:
    """Simple placeholder for progress tracking"""
    def __init__(self):
        pass

def show_memory_usage(label: str):
    """Simple memory usage display"""
    process = psutil.Process(os.getpid())
    memory_mb = process.memory_info().rss / 1024 / 1024
    print(f"{label}: {memory_mb:.2f} MB")

# Configuration for NZ Legislation Loophole Analysis Dataset Creation
INPUT_FILE = "nz-legislation.txt"  # Path to New Zealand legislation JSON dataset
OUTPUT_DIR = "nz_legislation_dataset"   # Directory to save the dataset
CHUNK_SIZE = 4096              # Size of text chunks for processing legislation sections
CHUNK_OVERLAP = 256            # Overlap between chunks to maintain context
BATCH_SIZE = 16                # Number of chunks to process at once
MODEL_PATH = "qwen3.gguf"  # Path to your Qwen3 GGUF model
MAX_TOKENS = 4096              # Maximum tokens for model response

# Ensure output directory exists
os.makedirs(OUTPUT_DIR, exist_ok=True)

def load_model(progress_manager: ProgressManager = None):
    """Load the LLM model for text generation with progress tracking"""
    if progress_manager is None:
        progress_manager = ProgressManager()

    print("Loading LLM model...")
    show_memory_usage("Initial memory usage")

    start_time = time.time()
    try:
        llm = Llama.from_pretrained(
            repo_id="DavidAU/Qwen3-Zero-Coder-Reasoning-0.8B-NEO-EX-GGUF",
            filename="Qwen3-Zero-Coder-Reasoning-0.8B-NEO-EX-D_AU-IQ4_XS-imat.gguf",
            n_ctx=40960,  # Context length
            n_threads=8,  # Adjust based on your CPU
            verbose=False,
            n_gpu_layers=-1,  # Use all available GPU layers
            n_batch=4096,  # Batch size for processing
            logits_all=False,  # Optimize for text generation
            use_mlock=True,  # Lock model in memory if possible
            use_mmap=True,  # Use memory mapping for better performance
        )
    except Exception as e:
        print(f"Error loading model: {e}")
        print("Trying with basic configuration...")
        # Fallback to basic configuration
        model = Llama(
            model_path=MODEL_PATH,
            n_ctx=40960,
            n_threads=8,
            verbose=False,
            n_gpu_layers=-1,
            n_batch=4096
        )

    load_time = time.time() - start_time
    print(f"LLM model loaded in {load_time:.2f}s")
    show_memory_usage("Memory after model load")

    return model

def clean_text(text: str) -> str:
    """Clean and normalize text for better embedding quality, optimized for legal/legislative content"""
    import re
    
    # Preserve section numbers and legal structure while cleaning
    # Keep section numbers like "1:", "2:", etc.
    text = re.sub(r'^(\d+:)', r'\1', text, flags=re.MULTILINE)
    
    # Remove excessive whitespace but preserve paragraph structure
    text = re.sub(r'[ \t]+', ' ', text)  # Replace multiple spaces/tabs with single space
    text = re.sub(r'\n\s*\n', '\n\n', text)  # Preserve paragraph breaks but clean up
    text = re.sub(r'\n{3,}', '\n\n', text)  # Reduce excessive newlines to double
    
    # Remove control characters but preserve legal formatting
    text = re.sub(r'[\x00-\x08\x0b\x0c\x0e-\x1f\x7f-\x9f]', '', text)  # Remove control chars except tab and newline
    
    # Handle legal-specific characters and formatting
    # Keep legal punctuation and symbols
    allowed_chars = r'\w\s\.\,\!\?\;\:\-\(\)\[\]\{\}\"\'\/\@\#\$\%\^\&\*\+\=\~\`°§'
    text = re.sub(r'[^' + allowed_chars + ']', '', text)
    
    # Normalize quotes and apostrophes for legal text
    text = re.sub(r'[""]', '"', text)  # Normalize double quotes
    text = re.sub(r"['']", "'", text)  # Normalize single quotes
    text = re.sub(r'`', "'", text)  # Replace backticks with apostrophes
    
    # Clean up legal numbering and references
    # Normalize section references
    text = re.sub(r'section\s+(\d+)', r'section \1', text, flags=re.IGNORECASE)
    text = re.sub(r'(\d+)\s*[Jj]anuary', r'\1 January', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Jj]uly', r'\1 July', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Aa]pril', r'\1 April', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Ff]ebruary', r'\1 February', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Dd]ecember', r'\1 December', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Aa]ugust', r'\1 August', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Mm]arch', r'\1 March', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Mm]ay', r'\1 May', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Jj]une', r'\1 June', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Ss]eptember', r'\1 September', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Oo]ctober', r'\1 October', text)  # Clean date formatting
    text = re.sub(r'(\d+)\s*[Nn]ovember', r'\1 November', text)  # Clean date formatting
    
    # Clean up punctuation spacing in legal text
    text = re.sub(r'\s+([\.!\?\,\;\:])', r'\1', text)  # Remove space before punctuation
    text = re.sub(r'([\.!\?\,\;\:])\s*', r'\1 ', text)  # Ensure space after punctuation
    
    # Handle legal citations and references (generic patterns)
    # Normalize act names with years - generic pattern for "Act ####" format
    text = re.sub(r'(\b\w+(?:\s+\w+)*)\s+Act\s+(\d{4})', r'\1 Act', text)  # Normalize act names
    
    # Clean up amendment references (generic patterns)
    text = re.sub(r'[Aa]mendment\(s\)\s+incorporated\s+in\s+the\s+[Aa]ct\(s\)', 'Amendments incorporated', text)
    text = re.sub(r'section\s+\d+\(\d+\)\([a-zA-Z]\)', lambda m: m.group(0).lower(), text)  # Normalize section references
    
    # Generic pattern for legal document sections
    text = re.sub(r'(\b(?:section|part|chapter|article|clause|subsection|paragraph))\s+(\d+[a-zA-Z]*)',
                  lambda m: f"{m.group(1)} {m.group(2)}", text, flags=re.IGNORECASE)

    # NZ-specific legal enhancements
    # Handle New Zealand specific terms and references
    text = re.sub(r'\b[Nn]ew\s+[Zz]ealand\b', 'New Zealand', text)  # Normalize "New Zealand"
    text = re.sub(r'\b[Pp]arliament\b', 'Parliament', text)  # Normalize "Parliament"
    text = re.sub(r'\b[Cc]rown\b', 'Crown', text)  # Normalize "Crown"
    text = re.sub(r'\b[Gg]overnment\b', 'Government', text)  # Normalize "Government"

    # Handle NZ-specific legal citations (e.g., "NZB" references, Treaty of Waitangi)
    text = re.sub(r'\b[Nn][Zz][Bb]\s+(\d+)', r'NZB \1', text)  # Normalize NZB references
    text = re.sub(r'[Tt]reaty\s+[Oo]f\s+[Ww]aitangi', 'Treaty of Waitangi', text, flags=re.IGNORECASE)

    # Handle Maori-specific characters if present (basic support)
    # Keep common Maori characters: ā, ē, ī, ō, ū, wh
    maori_chars = 'āēīōūwhĀĒĪŌŪWH'
    allowed_chars += maori_chars
    text = re.sub(r'[^' + allowed_chars + ']', '', text)
    
    # Remove empty lines and trim while preserving legal structure
    lines = []
    for line in text.split('\n'):
        stripped = line.strip()
        if stripped:  # Keep non-empty lines
            # Preserve section headers
            if re.match(r'^\d+:', stripped):
                lines.append(stripped)
            else:
                lines.append(stripped)
    
    text = '\n'.join(lines)
    
    # Final cleanup
    text = text.strip()
    
    return text

# Constants for prompt formatting
REASONING_START = "<start_working_out>"
REASONING_END = "<end_working_out>"
SOLUTION_START = "<SOLUTION>"
SOLUTION_END = "</SOLUTION>"

def create_system_prompt(text: str) -> str:
    """Create a system prompt for analyzing legislative text for loopholes and ambiguities"""
    return f"""You are a legal expert analyzing New Zealand legislation for loopholes and ambiguities.

LEGISLATION TEXT:
{text}

TASK: Analyze this legislative text and identify potential loopholes, ambiguities, or unintended consequences.

REASONING: Provide a structured analysis in the following format:

1. **Text Meaning**: Explain what the text means and its intended purpose
2. **Key Assumptions**: Identify any assumptions the text makes that could be exploited
3. **Exploitable Interpretations**: Discuss how the text could be interpreted or applied in ways that circumvent its intended purpose
4. **Critical Loopholes**: Identify specific loopholes, ambiguities, or unintended consequences that could be used to bypass the legislation
5. **Circumvention Strategies**: Suggest practical methods or scenarios for exploiting these loopholes to achieve objectives contrary to the legislation's intent

Write your complete analysis between {REASONING_START} and {REASONING_END}.

Then provide your overall conclusion between {SOLUTION_START} and {SOLUTION_END}.
"""

def generate_chat_template(system_prompt: str) -> str:
    """
    Generate a chat template using the GGUF model's native chat format.
    This uses the proper message structure with BOS/EOS tokens for better model compatibility.
    """
    # Build the chat using the GGUF template structure
    chat_messages = []

    # System message
    if system_prompt:
        chat_messages.append("<|im_start|>system")
        chat_messages.append(system_prompt)
        chat_messages.append("<|im_end|>")

    # User message with the analysis request
    chat_messages.append("<|im_start|>user")
    chat_messages.append("Analyze the given legislative text for loopholes, ambiguities, and unintended consequences. Provide a structured legal analysis following the specified format.")
    chat_messages.append("<|im_end|>")

    # Assistant message with generation prompt
    chat_messages.append("<|im_start|>assistant")
    chat_messages.append("")  # Empty for generation

    return "\n".join(chat_messages)

def chunk_text(text: str, chunk_size: int = CHUNK_SIZE, overlap: int = CHUNK_OVERLAP) -> List[str]:
    """Split text into overlapping chunks for processing"""
    if len(text) <= chunk_size:
        return [text]

    chunks = []
    start = 0
    while start < len(text):
        end = start + chunk_size
        chunk = text[start:end]

        # Try to end chunk at a sentence boundary if possible
        if end < len(text):
            # Look for sentence endings in the last 100 characters
            sentence_end = max(
                chunk.rfind('. ', max(0, len(chunk) - 100)),
                chunk.rfind('! ', max(0, len(chunk) - 100)),
                chunk.rfind('? ', max(0, len(chunk) - 100))
            )
            if sentence_end != -1:
                chunk = chunk[:sentence_end + 2]  # Include the sentence ending

        chunks.append(chunk)
        start = end - overlap if end < len(text) else len(text)

    return chunks

def generate_response(model, prompt: str, max_tokens: int = MAX_TOKENS) -> str:
    """
    Generate a response from the model for a given prompt with optimized parameters for legal analysis.

    Parameter Explanations:
    - temperature=0.3: Balanced creativity for legal analysis (not too random, not too deterministic)
    - top_p=0.85: Nucleus sampling - considers top 85% probability mass for coherent legal text
    - top_k=50: Top-k sampling - considers top 50 tokens for better legal terminology selection
    - min_p=0.05: Minimum probability threshold to avoid low-quality tokens

    Anti-Repetition Parameters:
    - repetition_penalty=1.15: Penalizes repetition of phrases (15% penalty)
    - presence_penalty=0.1: Encourages topic diversity across the response
    - frequency_penalty=0.1: Reduces overuse of frequent tokens

    Advanced Sampling:
    - typical_p=0.95: Focuses on typical token probabilities for legal text patterns
    - tfs_z=0.95: Tail-free sampling for more natural legal reasoning
    - mirostat_mode=2: Mirostat v2 for perplexity-controlled generation
    - mirostat_tau=4.0: Target entropy level for legal analysis
    - mirostat_eta=0.15: Learning rate for perplexity adaptation
    """
    try:
        response = model(
            prompt,
            max_tokens=max_tokens,
            # Core generation parameters
            temperature=0.3,  # Balanced temperature for legal analysis
            top_p=0.85,       # Nucleus sampling for coherent legal text
            top_k=50,         # Top-k sampling for better token selection
            min_p=0.05,       # Minimum probability threshold to avoid low-quality tokens

            # Anti-repetition parameters
            repeat_penalty=1.15,    # Reduce repetition of phrases
            presence_penalty=0.1,       # Encourage topic diversity
            frequency_penalty=0.1,      # Reduce frequent token usage

            # Advanced sampling parameters
            typical_p=0.95,    # Typical token probability for legal text patterns
            tfs_z=0.95,        # Tail-free sampling for better reasoning
            mirostat_mode=2,   # Mirostat v2 for perplexity control
            mirostat_tau=4.0,  # Mirostat target entropy
            mirostat_eta=0.15, # Mirostat learning rate

            # Stopping conditions
            stop=[SOLUTION_END, "</SOLUTION>", "<end_working_out>"]  # Multiple stop tokens
        )
        return response['choices'][0]['text'].strip()
    except Exception as e:
        print(f"Error generating response: {e}")
        # Try with fallback parameters if advanced ones fail
        try:
            response = model(
                prompt,
                max_tokens=max_tokens,
                temperature=0.3,
                top_p=0.85,
                top_k=50,
                repeat_penalty=1.15,
                stop=[SOLUTION_END, "</SOLUTION>"]
            )
            return response['choices'][0]['text'].strip()
        except Exception as e2:
            print(f"Fallback also failed: {e2}")
            return ""

def parse_legislation_json(file_path: str) -> List[Dict[str, Any]]:
    """Parse the JSON lines format of NZ legislation dataset"""
    legislation_entries = []

    try:
        with open(file_path, 'r', encoding='utf-8') as f:
            for line_num, line in enumerate(f, 1):
                line = line.strip()
                if line:
                    try:
                        entry = json.loads(line)
                        if 'id' in entry and 'text' in entry:
                            legislation_entries.append(entry)
                        else:
                            print(f"Warning: Line {line_num} missing required fields, skipping")
                    except json.JSONDecodeError as e:
                        print(f"Warning: Could not parse line {line_num}: {e}")
                        continue
    except Exception as e:
        print(f"Error reading legislation file: {e}")
        return []

    print(f"Successfully parsed {len(legislation_entries)} legislation entries")
    return legislation_entries

def create_finetuning_dataset(input_file: str, model, output_file: str = None) -> List[Dict[str, Any]]:
    """Create a finetuning dataset by processing NZ legislation JSON dataset with incremental saving"""
    if output_file is None:
        output_file = os.path.join(OUTPUT_DIR, "nz_legislation_loophole_dataset.json")

    # Create temporary file paths
    temp_file = output_file.replace('.json', '_temp.jsonl')
    backup_file = output_file.replace('.json', '_backup.json')

    print(f"Parsing legislation dataset from {input_file}")
    legislation_entries = parse_legislation_json(input_file)

    if not legislation_entries:
        print("No legislation entries found to process")
        return []

    dataset = []
    total_entries = len(legislation_entries)
    saved_count = 0

    print(f"Processing {total_entries} legislation entries...")
    print(f"Dataset will be saved incrementally to: {temp_file}")

    try:
        # Open temporary file for incremental saving
        with open(temp_file, 'w', encoding='utf-8') as temp_f:
            for entry_num, entry in enumerate(legislation_entries, 1):
                legislation_id = entry.get('id', f'entry_{entry_num}')
                title = entry.get('title', 'Unknown Title')
                year = entry.get('year', 'Unknown Year')
                raw_text = entry.get('text', '')

                print(f"\nProcessing entry {entry_num}/{total_entries}: {title} ({year}) - ID: {legislation_id}")

                # Clean the legislation text
                cleaned_text = clean_text(raw_text)

                # Chunk the text if it's too long
                chunks = chunk_text(cleaned_text)

                print(f"  - Text length: {len(raw_text)} characters")
                print(f"  - Number of chunks: {len(chunks)}")

                # Process each chunk
                for chunk_id, chunk in enumerate(chunks):
                    # Create prompt for this chunk
                    system_prompt = create_system_prompt(chunk)
                    full_prompt = generate_chat_template(system_prompt)

                    # Generate response
                    response = generate_response(model, full_prompt)

                    # Print response for monitoring
                    print(f"\n📝 **Generated Analysis for {title} (Chunk {chunk_id + 1}/{len(chunks)})**:")
                    print(f"   Response length: {len(response)} characters")

                    # Show preview of the analysis
                    preview = response.replace('\n', ' ').strip()
                    print(f"   Preview: {preview}")

                    # Check for key analysis elements
                    has_reasoning = '<start_working_out>' in response or 'reasoning' in response.lower()
                    has_loopholes = 'loophole' in response.lower() or 'ambiguity' in response.lower() or 'issue' in response.lower()
                    has_recommendations = 'recommend' in response.lower() or 'suggest' in response.lower()

                    print(f"   Analysis quality: {'✅' if has_reasoning else '❌'} Reasoning | {'✅' if has_loopholes else '❌'} Loopholes | {'✅' if has_recommendations else '❌'} Recommendations")

                    # Add to dataset with metadata
                    dataset_entry = {
                        "prompt": full_prompt,
                        "response": response,
                        "legislation_id": legislation_id,
                        "title": title,
                        "year": year,
                        "chunk_id": chunk_id,
                        "total_chunks": len(chunks),
                        "text_length": len(chunk),
                        "original_text_length": len(raw_text)
                    }

                    # Save entry immediately to temporary file (JSON Lines format)
                    json.dump(dataset_entry, temp_f, ensure_ascii=False)
                    temp_f.write('\n')
                    temp_f.flush()  # Force write to disk

                    dataset.append(dataset_entry)
                    saved_count += 1

                    # Progress update every 10 entries
                    if saved_count % 10 == 0:
                        print(f"  ✓ Saved {saved_count} entries so far...")

        print(f"\n✓ All entries processed and saved to temporary file")
        print(f"✓ Total entries saved: {saved_count}")

        # Create backup of existing file if it exists
        if os.path.exists(output_file):
            print(f"Creating backup of existing dataset...")
            os.rename(output_file, backup_file)

        # Convert JSON Lines to final JSON format
        print(f"Converting to final JSON format...")
        with open(temp_file, 'r', encoding='utf-8') as temp_f:
            lines = temp_f.readlines()

        final_dataset = []
        for line in lines:
            if line.strip():
                final_dataset.append(json.loads(line))

        # Save final consolidated JSON file
        with open(output_file, 'w', encoding='utf-8') as f:
            json.dump(final_dataset, f, indent=2, ensure_ascii=False)

        print(f"✓ Final dataset saved to: {output_file}")

        # Clean up temporary file
        if os.path.exists(temp_file):
            os.remove(temp_file)
            print(f"✓ Temporary file cleaned up")

        # Clean up backup file if everything succeeded
        if os.path.exists(backup_file):
            os.remove(backup_file)
            print(f"✓ Backup file cleaned up")

        print(f"\n🎉 Dataset creation complete!")
        print(f"   • Processed {total_entries} legislation documents")
        print(f"   • Generated {len(final_dataset)} analysis entries")
        print(f"   • Total chunks processed: {sum(entry.get('total_chunks', 1) for entry in final_dataset[:total_entries])}")

        return final_dataset

    except KeyboardInterrupt:
        print(f"\n⚠️  Process interrupted by user")
        print(f"   • Partial dataset saved to: {temp_file}")
        print(f"   • {saved_count} entries saved so far")
        print(f"   • You can resume processing or use the temporary file")
        raise

    except Exception as e:
        print(f"\n❌ Error during processing: {e}")
        print(f"   • Partial dataset saved to: {temp_file}")
        print(f"   • {saved_count} entries saved so far")
        if os.path.exists(backup_file):
            print(f"   • Original dataset restored from backup")
            os.rename(backup_file, output_file)
        raise

def main():
    """Main execution function"""
    print("Starting NZ Legislation Loophole Analysis Dataset Creation")
    print("=" * 60)

    # Load the model
    model = load_model()

    # Create the dataset
    dataset = create_finetuning_dataset(INPUT_FILE, model)

    # Cleanup
    if hasattr(model, 'close'):
        model.close()

    print("\nDataset creation completed successfully!")
    print(f"Output saved to: {os.path.join(OUTPUT_DIR, 'nz_legislation_loophole_dataset.json')}")

if __name__ == "__main__":
    main()