Personal Clone with WhatsApp Messages

A Python toolkit for creating a personalized AI clone of yourself using WhatsApp chat data. This tool analyzes your chat patterns, writing style, and personality traits to create a fine-tuned language model that can interact just like you.

Features

• Advanced Style Analysis:

  • Message length patterns
  • Emoji usage frequency
  • Slang and abbreviation patterns
  • Capitalization habits
  • Punctuation patterns
  • Common phrases and expressions
  • Response patterns

• Multiple LLM Support:

  • Llama-2
  • Llama-3
  • Mistral
  • Falcon
  • GPT
  • Qwen

• Optimization Features:

  • Parameter-Efficient Fine-Tuning (LoRA)
  • 8-bit quantization support
  • Gradient accumulation
  • Mixed precision training

Requirements

pip install pandas transformers torch datasets tensorboard peft

Getting Started

1. Chat Export and Parsing

Chat Format Requirements

Your WhatsApp chat export should follow this format:

[MM/DD/YY, HH:MM:SS AM/PM] Author: Message text here

Example:

[01/13/24, 12:24:48 AM] Alex: Have you finished that project for work?
[01/13/24, 12:52:48 AM] Jamie: I love those! Send me the title later.

Supported Date Formats

1. Standard WhatsApp format: [MM/DD/YY, HH:MM:SS AM/PM]
2. Without seconds: [MM/DD/YY, HH:MM AM/PM]
3. International format: [DD/MM/YY, HH:MM:SS AM/PM]
4. ISO-like format: [YYYY-MM-DD, HH:MM:SS AM/PM]

Export Instructions

1. Open WhatsApp chat
2. Tap ⋮ (three dots) > More > Export chat
3. Choose 'Without media'
4. Save the .txt file

What to Include/Exclude

• ✅ Include:

  • Regular text messages
  • Emoji messages
  • URLs (they will be cleaned automatically)
  • Normal conversation text

• ❌ Exclude:

  • Media messages (images, videos, documents)
  • System messages
  • Group settings changes
  • Contact cards
  • Location shares

Parsing Your Chat

Basic usage:

python parser.py chat.txt "YourName" "OtherPerson" "YourName"

Advanced usage:

python parser.py chat.txt "YourName" "OtherPerson" "YourName" \
    --llm_format mistral \
    --context_length 5

This generates:

• output_YYYYMMDD_HHMMSS.csv: Original conversation pairs
• formatted_YourName.jsonl: Training data
• style_metrics_YourName.json: Style analysis

2. Fine-tuning Process

After parsing your chat data, you can fine-tune a model:

python finetune.py \
    --data_path formatted_YourName.jsonl \
    --model_name "mistralai/Mistral-7B-v0.1" \
    --output_dir "./my_chatbot" \
    --style_metrics_path style_metrics_YourName.json \
    --use_peft \
    --use_8bit

Model Selection Guide

1. For Personal Use (Lower Resources):

   • Llama3-8B (16GB VRAM)
   • Qwen-7B (16GB VRAM)
   • Mistral 7B (16GB VRAM)
   • Llama-2 7B (16GB VRAM)

2. For Better Quality (Higher Resources):

   • Llama3-70B (80GB VRAM)
   • Qwen-14B (28GB VRAM)
   • Llama-2 13B (24GB VRAM)
   • Falcon 40B (Multiple GPUs)

How Style Analysis Works

The tool performs a comprehensive analysis of your chat style through multiple layers:

Message Structure Analysis

• Length Patterns

  • Average message length
  • Message length distribution
  • Typical response lengths

• Emoji Usage

  • Detection and frequency analysis
  • Favorite emoji patterns
  • Contextual emoji usage

• Slang and Abbreviations

  • Common internet slang (e.g., 'lol', 'omg', 'idk')
  • Personal abbreviations
  • Informal language patterns

Writing Style Metrics

• Capitalization

  • Sentence start patterns
  • Stylistic caps usage
  • Name/proper noun capitalization

• Punctuation

  • End-of-sentence patterns
  • Multiple punctuation usage (!!!, ???)
  • Informal punctuation style

• Common Phrases

  • Frequent expressions
  • Conversation starters/enders
  • Personal catchphrases

Advanced Training Options

Memory Optimization

1. 8-bit Quantization

   python finetune.py \
       --data_path formatted_YourName.jsonl \
       --use_8bit \
       --batch_size 2

2. Gradient Accumulation

   python finetune.py \
       --gradient_accumulation_steps 8 \
       --batch_size 1

Model-Specific Features

Llama3 Support

• BF16 precision training
• Flash Attention 2
• 8192 token context
• Example:

  python finetune.py \
      --model_name "meta-llama/Llama-3-8b" \
      --use_flash_attention \
      --bf16

Qwen Support

• Custom chat template
• Flash Attention
• Example:

  python finetune.py \
      --model_name "Qwen/Qwen-7B" \
      --use_flash_attention

Troubleshooting

Common Issues

1. Memory Errors

   • Reduce batch size
   • Enable 8-bit training
   • Use gradient accumulation

   python finetune.py \
       --batch_size 1 \
       --use_8bit \
       --gradient_accumulation_steps 8

2. Training Issues

   • Adjust learning rate
   • Increase training data
   • Try different models

   python finetune.py \
       --learning_rate 1e-5 \
       --num_epochs 5

3. Parsing Issues

   • Check date format
   • Verify chat export
   • Clean input data

Hardware Requirements

1. GPU Memory Requirements

   • 7B models: 16GB VRAM
   • 13B models: 24GB VRAM
   • 70B models: 80GB VRAM
   • Multiple GPUs: Falcon 40B

2. Recommended Setup

   • NVIDIA GPU with CUDA support
   • 32GB+ System RAM
   • SSD for faster data loading

Best Practices

Data Preparation

1. Use at least 1000 messages
2. Include diverse conversations
3. Clean irrelevant messages
4. Maintain conversation context

Training Configuration

1. Start with smaller models
2. Use default hyperparameters
3. Monitor with TensorBoard
4. Save checkpoints regularly

Style Preservation

1. Keep emoji patterns
2. Maintain punctuation style
3. Preserve message length patterns
4. Retain personal phrases

Known Limitations

1. Data Format

   • WhatsApp format changes
   • Regional date formats
   • Media message handling

2. Resource Requirements

   • High VRAM usage
   • Long training times
   • System RAM needs

3. Model Access

   • Llama model approval
   • Usage restrictions
   • License requirements

4. Quality Factors

   • Short conversation impact
   • Mixed language handling
   • Group chat dynamics

Contributing

Feel free to:

• Submit issues
• Propose features
• Share improvements
• Report bugs

License

This project is licensed under the MIT License. See LICENSE file for details.

Acknowledgments

• HuggingFace Transformers
• Meta AI (Llama models)
• WhatsApp chat format documentation