Rankers

A lightweight, flexible library for training neural retrievers with HuggingFace transformers — featuring multiple ranking architectures, integrated PyTerrier support, and production-ready evaluation pipelines.

---

📘 Overview

Rankers provides a unified interface for training, evaluating, and deploying neural ranking models. Built on top of HuggingFace transformers, it supports multiple architectures (bi-encoders, cross-encoders, sparse retrievers, and more) while maintaining compatibility with the transformers.Trainer API.

• Multiple Architectures: Bi-encoders (Dot), Cross-encoders (CAT), Sequence-to-Sequence, Sparse models, BGE, and more
• HuggingFace Compatible: Drop-in RankerTrainer that extends transformers.Trainer
• PyTerrier Integration: Convert trained models to PyTerrier pipelines instantly
• Production Ready: Built-in evaluation, checkpointing, and loss functions optimized for ranking

---

🚀 Getting Started

1. Installation

Clone the repository and install dependencies:

git clone https://github.com/Parry-Parry/rankers.git
cd rankers
pip install -e .

Or install the latest from PyPI:

pip install rankers

2. Quick Start: Train a Bi-Encoder

from rankers import RankerTrainer, RankerTrainingArguments
from rankers.modelling import Dot
from rankers.datasets import TrainingDataset, Corpus

# Load pre-trained model
model = Dot.from_pretrained("bert-base-uncased")

# Prepare datasets
corpus = Corpus.from_ir_datasets("msmarco-passage")
train_dataset = TrainingDataset("train.jsonl", corpus=corpus, group_size=4)
eval_dataset = EvaluationDataset.from_qrels("qrels.txt", corpus=corpus)

# Configure training
args = RankerTrainingArguments(
    output_dir="./output",
    num_train_epochs=3,
    per_device_train_batch_size=32,
    eval_strategy="epoch",
    save_strategy="best",
)

# Train
trainer = RankerTrainer(
    model=model,
    args=args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    loss_fn="margin_mse",
)
trainer.train()

# Convert to PyTerrier
ranker = model.to_pyterrier()
results = ranker.transform(test_queries)

---

🧩 Supported Models

Rankers includes implementations of popular ranking architectures:

Each model supports:

• Training with RankerTrainer
• Evaluation with IR metrics (nDCG, MRR, MAP, etc.)
• PyTerrier conversion for deployment
• Checkpointing with save_pretrained() / from_pretrained()

---

📊 Data Format

Training Data (JSONL)

{"query_id": "q1", "doc_id_a": "d1", "doc_id_b": "d2"}
{"query_id": "q2", "doc_id_a": "d3", "doc_id_b": "d4"}

Evaluation Data (Qrels)

q1 0 d1 2
q1 0 d2 1
q2 0 d3 2

---

⚙️ Training Configuration

Customize your training with RankerTrainingArguments:

from rankers import RankerTrainingArguments

args = RankerTrainingArguments(
    output_dir="./checkpoints",
    num_train_epochs=3,
    per_device_train_batch_size=32,
    per_device_eval_batch_size=64,
    learning_rate=2e-5,
    warmup_steps=500,

    # Evaluation & Checkpointing
    eval_strategy="steps",
    eval_steps=500,
    save_strategy="best",
    metric_for_best_model="eval_nDCG@10",
    load_best_model_at_end=True,

    # Loss & Regularization
    loss_fn="margin_mse",  # or "contrastive", "listwise", etc.
    regularizer="l2",
    regularization_weight=0.001,

    # Tracking
    report_to=["wandb"],
)

---

🎯 Loss Functions

Rankers includes multiple loss functions optimized for ranking:

• Pairwise: margin_mse, contrastive
• Listwise: listnet, ndcg_loss
• Ranking-specific: triplet, in_batch_negatives

---

🔄 Evaluation Pipeline

Built-in evaluation with IR metrics:

# During training (automatic)
trainer = RankerTrainer(
    ...,
    eval_dataset=eval_dataset,
)
results = trainer.evaluate()
# Returns: {"eval_nDCG@10": 0.45, "eval_MRR": 0.52, ...}

# After training
predictions = trainer.predict(test_dataset)
# Returns: PredictionOutput with scores and metrics

Supported metrics (via ir_measures):

• nDCG@k, MRR, MAP, Recall@k, Precision@k, and more

---

🚀 Deployment: PyTerrier Integration

Once trained, convert to PyTerrier for deployment:

# Load trained model
model = Dot.from_pretrained("./checkpoints/best_model")

# Convert to PyTerrier
ranker = model.to_pyterrier(batch_size=128)

# Use in IR pipelines
pipeline = retriever >> ranker >> reranker

# Run end-to-end ranking
results = pipeline.transform(queries_df)

---

📖 Examples

Check the examples/ directory for complete scripts:

• train.bert.cat.py: Train a BERT-based cross-encoder
• train.biencoder.py: Train a bi-encoder with in-batch negatives
• train.sparse.py: Train a sparse neural retriever
• eval_and_rank.py: Evaluate and generate rankings

Run an example:

python examples/train.bert.cat.py \
    --model_name_or_path bert-base-uncased \
    --training_data path/to/train.jsonl \
    --output_dir ./my_ranker

---

📚 Documentation

Full API documentation is available and can be built using Sphinx:

# Install documentation dependencies
pip install -e ".[docs]"

# Build documentation
cd docs && make html

# View in browser
open _build/html/index.html

The documentation includes:

• Complete API reference for all modules
• Architecture deep-dives and hyperparameter guides
• Training tutorials and best practices
• PyTerrier integration examples

---

🧪 Testing

Comprehensive test suite with integration tests:

# Run all tests
pytest

# Run only integration tests
pytest tests/integration/

# Run with coverage
pytest --cov=rankers

All tests pass ✅ including:

• 89 integration & unit tests covering full training pipeline
• Gradient flow verification across evaluation passes
• Model checkpointing and loading functionality
• Loss computation and backward propagation

---

🤝 Contributing

We welcome contributions! Please read our contributing guidelines for more details.

---

👥 Authors

• Andrew Parry

---

📄 License

This project is licensed under the Apache 2.0 License — see the LICENSE file for details.

---

🔗 Related Projects

• PyTerrier — Information Retrieval research platform
• HuggingFace Transformers — State-of-the-art NLP models
• ir_measures — Standard IR evaluation metrics