Phinetuning

Advanced finetuning toolkit for Phi-2 and other language models, featuring distributed training, hyperparameter optimization, and mixed precision training.

Table of Contents

• Features
• Installation
• Usage
  • Training
  • Generation
• Advanced Configuration
• Examples
• Contributing
• License

Features

• Distributed training support
• Mixed precision training (FP16/BF16)
• 4-bit and 8-bit quantization
• Hyperparameter optimization using Optuna
• Data augmentation techniques
• Advanced text generation with various decoding strategies
• Streaming generation support
• Comprehensive logging and error handling
• Support for multiple model architectures

Installation

git clone https://github.com/murapa96/Phinetuning.git
cd Phinetuning
pip install -r requirements.txt

Usage

Training

Basic training command:

python train.py \
  --model_path microsoft/phi-2 \
  --dataset_path your_dataset.json \
  --output_dir ./results \
  --batch_size 1 \
  --grad_accum_steps 4

Enable advanced features:

python train.py \
  --model_path microsoft/phi-2 \
  --dataset_path your_dataset.json \
  --output_dir ./results \
  --mixed_precision bf16 \
  --load_in_4bit \
  --distributed \
  --tune_hyperparams \
  --n_trials 10 \
  --augment_data

Generation

Basic text generation:

python generate.py \
  --model_path ./results/final_model \
  --input_text "Your prompt here" \
  --max_length 200

Advanced generation with sampling:

python generate.py \
  --model_path ./results/final_model \
  --input_file inputs.json \
  --output_file outputs.json \
  --temperature 0.7 \
  --top_p 0.95 \
  --top_k 50 \
  --do_sample \
  --streaming \
  --mixed_precision

Advanced Configuration

Training Arguments

• --mixed_precision: Choose between 'no', 'fp16', or 'bf16'
• --load_in_4bit: Enable 4-bit quantization
• --load_in_8bit: Enable 8-bit quantization
• --distributed: Enable distributed training
• --tune_hyperparams: Enable Optuna hyperparameter tuning
• --augment_data: Enable data augmentation
• --max_seq_length: Maximum sequence length (default: 2048)
• --learning_rate: Learning rate (default: 2e-4)
• --batch_size: Per device batch size
• --grad_accum_steps: Gradient accumulation steps

Generation Arguments

• --temperature: Controls randomness (0.0-1.0)
• --top_k: Top-k sampling parameter
• --top_p: Nucleus sampling parameter
• --num_beams: Number of beams for beam search
• --repetition_penalty: Penalize repeated tokens
• --streaming: Enable token-by-token generation
• --mixed_precision: Enable mixed precision inference
• --load_in_4bit: Enable 4-bit quantization for memory efficiency

Examples

1. Distributed training with mixed precision:

python train.py \
  --model_path microsoft/phi-2 \
  --dataset_path your_dataset.json \
  --distributed \
  --mixed_precision bf16 \
  --batch_size 2 \
  --grad_accum_steps 4 \
  --learning_rate 2e-4 \
  --max_steps 10000

2. Hyperparameter optimization:

python train.py \
  --model_path microsoft/phi-2 \
  --dataset_path your_dataset.json \
  --tune_hyperparams \
  --n_trials 20 \
  --load_in_4bit

3. Batch generation with advanced sampling:

python generate.py \
  --model_path ./results/final_model \
  --input_file inputs.json \
  --output_file outputs.json \
  --batch_size 4 \
  --temperature 0.7 \
  --top_p 0.95 \
  --do_sample \
  --mixed_precision

Contributing

We welcome contributions! Please:

1. Fork the repository
2. Create a feature branch: git checkout -b feature-name
3. Commit your changes: git commit -am 'Add feature'
4. Push to the branch: git push origin feature-name
5. Submit a pull request

License

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