unbabel-comet

What it is and what it does

unbabel-comet is a neural machine translation evaluation framework that scores translation quality using deep learning models trained on human assessments. It offers three evaluation modes: reference-based scoring (comparing translations to gold references), reference-free scoring (assessing translations without references), and explainable scoring (XCOMET models that identify error spans and classify them by severity according to MQM typology). The package includes CLI tools for batch scoring, system comparison with statistical significance testing, and Minimum Bayes Risk decoding for selecting best translations from candidate lists.

The framework depends on PyTorch, Hugging Face Transformers, and PyTorch Lightning, making it suitable for research and production environments where neural metric evaluation is needed. It supports document-level context through DocCOMET, enabling better handling of discourse phenomena and chat translation. Multiple pre-trained models are available via Hugging Face Hub, ranging from lightweight reference-free variants to large explainable models with billions of parameters.

Use it for:

• Evaluate machine translation system outputs during development and benchmarking against reference translations
• Score translations without references using quality-estimation models for deployment scenarios where gold references are unavailable
• Identify and classify specific translation errors (minor, major, critical) using XCOMET models for error analysis and system improvement
• Compare multiple MT systems with statistical significance testing to determine which performs best
• Rank and select best translations from candidate pools using Minimum Bayes Risk decoding for quality-aware decoding
• Evaluate document-level translation quality using context-aware DocCOMET models for improved accuracy on discourse phenomena

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Evaluates machine translation quality using neural metrics, scoring translations against references or without references, and explaining error types according to MQM typology.

Yes. Active maintenance, permissive Apache-2.0 license, low install friction, no known vulnerabilities, and substantial community adoption (774 stars, top 15000 tier). The package is production-ready for MT evaluation workflows. Install if you need neural metric evaluation; skip if you require only lightweight statistical metrics like BLEU.

Install

unbabel-comet on PyPI

pip install unbabel-comet

uv add unbabel-comet

poetry add unbabel-comet

Installing unbabel-comet

Before you install

Low friction installation via wheel distribution. Active maintenance with recent commits; 774 repository stars. Depends on 13 runtime packages including torch, transformers, and pytorch-lightning, which are substantial but standard for ML workflows.

License in practice

Apache-2.0 permissive license allows commercial and private use with minimal restrictions; suitable for most deployment scenarios.

Quickstart

pip install unbabel-comet

comet-score -s src.txt -t hyp.txt -r ref.txt

# Or in Python:
from comet import load_from_checkpoint
model = load_from_checkpoint("Unbabel/wmt22-comet-da")
scores = model.predict(data, batch_size=8)

Requires Python 3.8 or above. Some models (e.g., Unbabel/wmt22-cometkiwi-da) require Hugging Face Hub authentication and license acknowledgment.

Verify before relying

• Whether GPU acceleration is required or CPU-only scoring is practical for typical workloads
• Specific accuracy/correlation metrics compared to other MT evaluation frameworks
• Whether document-level context (DocCOMET) improves results for your language pair and domain

Package facts

Evidence: unbabel_comet-2.2.7-py3-none-any.whl

Keywords: Machine Translation, Evaluation, Unbabel, COMET

Tags

More Artificial Intelligence packages

Further reading