Qwen Fine-tuning on TL;DR Dataset

This project provides a complete pipeline for fine-tuning a small Qwen model on the TL;DR (Too Long; Didn't Read) dataset for text summarization tasks.

Features

• Small Model Support: Uses Qwen1.5-0.5B (500M parameters) for efficient training
• TL;DR Dataset: Fine-tuned on the popular TL;DR dataset from HuggingFace
• Easy-to-use API: Simple interface for training and inference
• Flexible Configuration: Customizable training parameters
• Memory Efficient: Optimized for training on consumer hardware

Installation

1. Install dependencies:

pip install -r requirements.txt

2. Verify installation:

python -c "import torch; print(f'PyTorch version: {torch.__version__}')"
python -c "import transformers; print(f'Transformers version: {transformers.__version__}')"

Quick Start

1. Test the Base Model

First, test the base model without fine-tuning:

python train.py --test_only

This will load the base Qwen model and generate a summary for a sample text.

2. Start Fine-tuning

Begin the fine-tuning process:

python train.py --num_epochs 3 --batch_size 4 --learning_rate 5e-5

3. Use the Fine-tuned Model

After training, test the fine-tuned model:

python train.py --load_model ./qwen-tldr-finetuned --test_only

Usage Examples

Basic Usage

from models import QwenSftModel

# Initialize model
model = QwenSftModel()

# Generate summary
text = "Your long text here..."
summary = model.generate_summary(text)
print(summary)

Custom Training

from models import QwenSftModel

# Initialize with custom model
model = QwenSftModel(model_name="Qwen/Qwen1.5-1B")

# Custom training parameters
trainer = model.train(
    output_dir="./my-finetuned-model",
    num_epochs=5,
    batch_size=2,
    learning_rate=3e-5
)

Load Fine-tuned Model

from models import QwenSftModel

# Load your fine-tuned model
model = QwenSftModel()
model.load_finetuned_model("./qwen-tldr-finetuned")

# Use for inference
summary = model.generate_summary("Your text here...")

Command Line Options

The train.py script supports various command-line arguments:

• --model_name: HuggingFace model name (default: "Qwen/Qwen1.5-0.5B")
• --output_dir: Directory to save the fine-tuned model (default: "./qwen-tldr-finetuned")
• --num_epochs: Number of training epochs (default: 3)
• --batch_size: Training batch size (default: 4)
• --learning_rate: Learning rate (default: 5e-5)
• --test_only: Only test the model without training
• --load_model: Path to load a pre-trained model

Model Architecture

The system uses:

• Base Model: Qwen1.5-0.5B (500M parameters)
• Dataset: TL;DR dataset from HuggingFace
• Training: Supervised fine-tuning with causal language modeling
• Prompt Format: "Summarize the following text:\n\n{text}\n\nSummary:"

Hardware Requirements

Minimum Requirements

• RAM: 8GB
• GPU: Not required (CPU training supported)
• Storage: 2GB free space

Recommended Requirements

• RAM: 16GB+
• GPU: NVIDIA GPU with 8GB+ VRAM
• Storage: 5GB free space

Training Configuration

The default training configuration is optimized for consumer hardware:

• Batch Size: 4 (adjust based on your GPU memory)
• Learning Rate: 5e-5 (good starting point for fine-tuning)
• Epochs: 3 (sufficient for most summarization tasks)
• Gradient Accumulation: 4 (effective batch size = 16)

Dataset Information

The TL;DR dataset contains:

• Train: ~300K examples
• Validation: ~30K examples
• Format: (text, summary) pairs
• Source: Reddit posts and their summaries

Troubleshooting

Common Issues

1. Out of Memory Error:

   • Reduce batch size: --batch_size 2
   • Use gradient accumulation: increase gradient_accumulation_steps

2. Slow Training:

   • Use GPU if available
   • Reduce dataset size for faster iteration

3. Poor Quality Summaries:

   • Increase training epochs
   • Adjust learning rate
   • Use larger base model

Performance Tips

• Use mixed precision training (enabled by default on GPU)
• Monitor GPU memory usage during training
• Start with smaller dataset subset for testing

Model Output

The fine-tuned model generates concise summaries in response to the prompt format:

Summarize the following text:

[Your input text]

Summary: [Generated summary]

License

This project uses the Qwen model which is subject to its own license. Please refer to the HuggingFace model page for licensing information.

Contributing

Feel free to submit issues and enhancement requests!