MPDistil 🎓

Meta-Policy Knowledge Distillation for Training Compact Student Models

MPDistil is a teacher-student collaborative knowledge distillation framework that enables compact student models to outperform larger teacher models through meta-learning and curriculum learning.

Based on the ICLR 2024 paper: A Good Learner can Teach Better: Teacher-Student Collaborative Knowledge Distillation

🌟 Key Features

• 📊 Superior Performance: 6-layer BERT student outperforms 12-layer BERT teacher on 5/6 SuperGLUE tasks
• 🎯 4-Phase Training: Teacher fine-tuning → PKD → Meta-teacher → Curriculum learning
• 🚀 Simple API: Easy-to-use .train() method with full control over all phases
• 📏 Flexible Metrics: Built-in support for accuracy, F1, MCC, correlation via HuggingFace evaluate
• 🔧 Customizable: Works with any HuggingFace model and custom datasets
• 💻 Colab-Ready: GPU-optimized for Google Colab environments
• 📦 Easy Installation: Single pip command to get started

📈 Methodology

MPDistil consists of 4 training phases:

1. Teacher Fine-tuning: Fine-tune teacher model on the target task
2. Student PKD: Knowledge distillation with Patient Knowledge Distillation
3. Meta-Teacher Learning: Collaborative or competitive loss for meta-learning
4. Curriculum Learning: Reinforcement learning-based task selection

🚀 Installation

From GitHub (Recommended)

pip install git+https://github.com/parmanu-lcs2/mpdistil.git

From Source

git clone https://github.com/parmanu-lcs2/mpdistil.git
cd mpdistil
pip install -e .

From pypi

pip install mpdistil

💡 Quick Start

Basic Usage

from mpdistil import MPDistil, load_superglue_dataset

# Load data
loaders, num_labels = load_superglue_dataset('CB')

# Initialize MPDistil
model = MPDistil(
    task_name='CB',
    num_labels=num_labels,
    metric='f1',  # Options: 'accuracy', 'f1', 'mcc', 'correlation', 'auto'
    teacher_model='bert-base-uncased',
    student_model='bert-base-uncased',
    student_layers=6
)

# Train with all 4 phases
history = model.train(
    train_loader=loaders['train'],
    val_loader=loaders['val'],
    teacher_epochs=10,   # Phase 1
    student_epochs=10,   # Phase 2
    meta_epochs=1        # Phase 3 (NEW!)
)

# Save trained student
model.save_student('./my_student_model')

# Make predictions
predictions = model.predict(loaders['test'])

With Custom Data

from mpdistil import MPDistil, create_simple_dataloader

# Prepare your data
texts = [("This is text A", "This is text B"), ...]
labels = [0, 1, 0, ...]

# Create DataLoader
train_loader = create_simple_dataloader(
    texts=texts,
    labels=labels,
    tokenizer_name='bert-base-uncased',
    max_length=128,
    batch_size=8
)

# Train model
model = MPDistil(task_name='MyTask', num_labels=2, metric='accuracy')
history = model.train(train_loader, val_loader)

With Meta-Learning (Curriculum)

# Load multiple tasks for curriculum learning
cb_loaders, _ = load_superglue_dataset('CB')
rte_loaders, _ = load_superglue_dataset('RTE')
boolq_loaders, _ = load_superglue_dataset('BoolQ')

# Train with curriculum learning
history = model.train(
    train_loader=cb_loaders['train'],
    val_loader=cb_loaders['val'],
    meta_loaders={
        'RTE': rte_loaders['val'],
        'BoolQ': boolq_loaders['val']
    },
    teacher_epochs=10,   # Phase 1
    student_epochs=10,   # Phase 2  
    meta_epochs=3,       # Phase 3 - can train for multiple epochs!
    num_episodes=200     # Phase 4 - curriculum learning episodes
)

📖 API Reference

MPDistil Class

Constructor

MPDistil(
    task_name: str,              # Name of the main task
    num_labels: int,             # Number of output classes
    metric: str = 'accuracy',    # Metric: 'accuracy', 'f1', 'mcc', 'correlation', 'auto'
    teacher_model: str = 'bert-base-uncased',  # HuggingFace model name
    student_model: str = 'bert-base-uncased',  # HuggingFace model name
    student_layers: int = 6,     # Number of layers for student
    device: str = 'auto',        # 'auto', 'cuda', or 'cpu'
    output_dir: str = './mpdistil_outputs'
)

Methods

Method	Description
train(train_loader, val_loader, **kwargs)	Train the model (all 4 phases)
predict(test_loader)	Generate predictions
save_student(path)	Save student model in HuggingFace format
load_student(path)	Load a saved student model
save_predictions(predictions, path, label_mapping)	Save predictions to TSV

TrainingConfig

Configure training hyperparameters:

from mpdistil import TrainingConfig

config = TrainingConfig(
    # Phase 1: Teacher
    teacher_epochs=10,
    teacher_lr=2e-5,
    
    # Phase 2: Student PKD
    student_epochs=10,
    student_lr=3e-5,
    alpha=0.5,          # Soft loss weight
    beta=100.0,         # PKD loss weight
    temperature=5.0,    # Distillation temperature
    
    # Phase 3: Meta-Teacher
    meta_epochs=3,      # NEW! Meta-teacher can train for multiple epochs
    meta_lr=1e-3,
    use_competitive_loss=False,  # Use collaborative loss
    
    # Phase 4: Curriculum
    num_episodes=200,
    reward_type='binary',  # or 'real'
    
    # General
    batch_size=8,
    seed=42,
    report_to=None      # Options: 'wandb', 'tensorboard', None
)

history = model.train(train_loader, val_loader, config=config)

Training Parameters

Parameter	Type	Default	Description
teacher_epochs	int	10	Phase 1: Teacher training epochs
student_epochs	int	10	Phase 2: Student training epochs
meta_epochs	int	1	Phase 3: Meta-teacher training epochs (NEW!)
num_episodes	int	200	Phase 4: Curriculum learning episodes
teacher_lr	float	2e-5	Teacher learning rate
student_lr	float	3e-5	Student learning rate
meta_lr	float	1e-3	Meta-learning rate
alpha	float	0.5	Soft vs hard loss weight
beta	float	100.0	PKD loss weight
temperature	float	5.0	Distillation temperature
use_competitive_loss	bool	False	Competitive vs collaborative
reward_type	str	'binary'	'binary' or 'real'
batch_size	int	8	Batch size
seed	int	42	Random seed

📊 Results

Performance on SuperGLUE tasks (BERT-base teacher → BERT-6L student):

Model	BoolQ	CB	COPA	RTE	WiC	WSC
Teacher	75.3	83.9	63.0	67.1	57.1	64.4
Student (Undistilled)	71.6	75.0	53.0	64.6	56.0	63.5
MPDistil (Ours)	73.4	83.9	70.0	67.5	59.6	65.4

✨ Student outperforms teacher on 5/6 tasks!

📝 Examples

See the examples/ directory for Jupyter notebooks:

• GLUE.ipynb: Usage with SuperGLUE

📊 Evaluation Metrics

MPDistil supports multiple evaluation metrics via HuggingFace evaluate library:

Available Metrics

Metric	Use Case	Returns
'accuracy'	Standard classification (default)	{'acc': float}
'f1'	Imbalanced datasets, multi-class	{'acc': float, 'f1': float, 'acc_and_f1': float}
'mcc'	Binary classification, imbalanced	{'mcc': float}
'correlation'	Regression tasks (STS-B)	{'pearson': float, 'spearmanr': float}
'auto'	Auto-detect based on task	Task-specific metric

Example: Using Different Metrics

# Accuracy (default)
model = MPDistil(task_name='BoolQ', num_labels=2, metric='accuracy')

# F1 score (recommended for CB, MultiRC)
model = MPDistil(task_name='CB', num_labels=3, metric='f1')

# Matthews Correlation (for binary with imbalance)
model = MPDistil(task_name='CoLA', num_labels=2, metric='mcc')

# Correlation (for regression)
model = MPDistil(task_name='STS-B', num_labels=1, metric='correlation', is_regression=True)

# Auto-detect
model = MPDistil(task_name='CB', num_labels=3, metric='auto')  # Uses F1 for CB

🔬 How It Works

Phase 1: Teacher Fine-tuning

Fine-tune a large teacher model (e.g., BERT-base) on your target task.

Phase 2: Student PKD

Train a smaller student model using:

• Soft targets from teacher (KL divergence)
• Hard labels (cross-entropy)
• Patient KD (intermediate feature matching)

Loss: α * soft_loss + (1-α) * hard_loss + β * pkd_loss

Phase 3: Meta-Teacher

Create a meta-teacher that learns from both teacher and student representations:

Collaborative loss (default):

L = 0.5 * CE(T'(h_teacher), y) + 0.5 * CE(T'(h_student), y)

Competitive loss:

L = -mean(P_teacher) + mean(P_student) + CE_loss

Phase 4: Curriculum Learning

Use reinforcement learning to select which auxiliary tasks help the student learn:

• Action model selects next task
• Reward based on student improvement over teacher
• REINFORCE algorithm updates policy

🛠️ Advanced Usage

Custom Model Architectures

model = MPDistil(
    task_name='MyTask',
    num_labels=3,
    teacher_model='roberta-large',
    student_model='distilbert-base-uncased',
    student_layers=6
)

Weights & Biases Logging

history = model.train(
    train_loader=train_loader,
    val_loader=val_loader,
    report_to='wandb',      # Options: 'wandb', 'tensorboard', None
    wandb_project='my-project',
    wandb_run_name='experiment-1'
)

Access Trained Models

# Access student model
student = model.student

# Access teacher model
teacher = model.teacher

# Use with HuggingFace
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained('./my_student_model')
model = AutoModel.from_pretrained('./my_student_model')

📚 Citation

If you use MPDistil in your research, please cite:

@inproceedings{sengupta2024mpdistil,
  title={A Good Learner can Teach Better: Teacher-Student Collaborative Knowledge Distillation},
  author={Sengupta, Ayan and Dixit, Shantanu and Akhtar, Md Shad and Chakraborty, Tanmoy},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://openreview.net/forum?id=Ixi4j6LtdX}
}

📄 License

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

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

🙏 Acknowledgments

• Original paper: ICLR 2024
• Built with PyTorch and HuggingFace Transformers

📧 Contact

For questions or issues, please open an issue on GitHub or contact the authors.

---

Made with ❤️ for the research community