Version 1.1 is out 🥳! whats new?

1. Updated documentation
2. Updated Pytorch Lightning version to avoid security vulnerabilities (Untrusted Data & Code Injection)
3. Inspired by Amrhein et al, 2022 we added the comet-mbr command for fast Minimum Bayes Risk Decoding.
4. New encoder models

Simple installation from PyPI

pip install --upgrade pip  # ensures that pip is current 
pip install unbabel-comet

or

pip install unbabel-comet==1.1.0 --use-feature=2020-resolver

To develop locally install Poetry and run the following commands:

git clone https://github.com/Unbabel/COMET
cd COMET
poetry install

Alternately, for development, you can run the CLI tools directly, e.g.,

PYTHONPATH=. ./comet/cli/score.py

Test examples:

echo -e "Dem Feuer konnte Einhalt geboten werden\nSchulen und Kindergärten wurden eröffnet." >> src.de
echo -e "The fire could be stopped\nSchools and kindergartens were open" >> hyp1.en
echo -e "The fire could have been stopped\nSchools and pre-school were open" >> hyp2.en
echo -e "They were able to control the fire.\nSchools and kindergartens opened" >> ref.en

Basic scoring command:

comet-score -s src.de -t hyp1.en -r ref.en

you can set --gpus 0 to test on CPU.

Scoring multiple systems:

comet-score -s src.de -t hyp1.en hyp2.en -r ref.en

WMT test sets via SacreBLEU:

comet-score -d wmt20:en-de -t PATH/TO/TRANSLATIONS

The default setting of comet-score prints the score for each segment individually. If you are only interested in the score for the whole dataset (computed as the average of the segment scores), you can use the --quiet flag.

comet-score -s src.de -t hyp1.en -r ref.en --quiet

You can select another model/metric with the --model flag and for reference-free (QE-as-a-metric) models you don't need to pass a reference.

comet-score -s src.de -t hyp1.en --model wmt21-comet-qe-mqm

Following the work on Uncertainty-Aware MT Evaluation you can use the --mc_dropout flag to get a variance/uncertainty value for each segment score. If this value is high, it means that the metric is less confident in that prediction.

comet-score -s src.de -t hyp1.en -r ref.en --mc_dropout 30

When comparing two MT systems we encourage you to run the comet-compare command to get statistical significance with Paired T-Test and bootstrap resampling (Koehn, et al 2004).

comet-compare -s src.de -x hyp1.en -y hyp2.en -r ref.en

New: Minimum Bayes Risk Decoding:

Inspired by Amrhein et al, 2022 work, we have developed a command to perform Minimum Bayes Risk decoding. This command receives a text file with source sentences and a text file containing all the MT samples and writes to an output file the best sample according to COMET.

comet-mbr -s [SOURCE].txt -t [MT_SAMPLES].txt --num_sample [X] -o [OUTPUT_FILE].txt

COMET is optimized to be used in a single GPU by taking advantage of length batching and embedding caching. When using Multi-GPU since data e spread across GPUs we will typically get fewer cache hits and the length batching samples is replaced by a DistributedSampler. Because of that, according to our experiments, using 1 GPU is faster than using 2 GPUs specially when scoring multiple systems for the same source and reference.

Nonetheless, if your data does not have repetitions and you have more than 1 GPU available, you can run multi-GPU inference with the following command:

comet-score -s src.de -t hyp1.en -r ref.en --gpus 2 --quiet

Warning: Segment-level scores using multigpu will be out of order. This is only useful for system scoring.

You can change the cache size of COMET using the following env variable:

export COMET_EMBEDDINGS_CACHE="2048"

by default the COMET cache size is 1024.

from comet import download_model, load_from_checkpoint

model_path = download_model("wmt20-comet-da")
model = load_from_checkpoint(model_path)
data = [
    {
        "src": "Dem Feuer konnte Einhalt geboten werden",
        "mt": "The fire could be stopped",
        "ref": "They were able to control the fire."
    },
    {
        "src": "Schulen und Kindergärten wurden eröffnet.",
        "mt": "Schools and kindergartens were open",
        "ref": "Schools and kindergartens opened"
    }
]
seg_scores, sys_score = model.predict(data, batch_size=8, gpus=1)

All the above mentioned models are build on top of XLM-R which cover the following languages:

Afrikaans, Albanian, Amharic, Arabic, Armenian, Assamese, Azerbaijani, Basque, Belarusian, Bengali, Bengali Romanized, Bosnian, Breton, Bulgarian, Burmese, Burmese, Catalan, Chinese (Simplified), Chinese (Traditional), Croatian, Czech, Danish, Dutch, English, Esperanto, Estonian, Filipino, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Hausa, Hebrew, Hindi, Hindi Romanized, Hungarian, Icelandic, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kazakh, Khmer, Korean, Kurdish (Kurmanji), Kyrgyz, Lao, Latin, Latvian, Lithuanian, Macedonian, Malagasy, Malay, Malayalam, Marathi, Mongolian, Nepali, Norwegian, Oriya, Oromo, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Sanskri, Scottish, Gaelic, Serbian, Sindhi, Sinhala, Slovak, Slovenian, Somali, Spanish, Sundanese, Swahili, Swedish, Tamil, Tamil Romanized, Telugu, Telugu Romanized, Thai, Turkish, Ukrainian, Urdu, Urdu Romanized, Uyghur, Uzbek, Vietnamese, Welsh, Western, Frisian, Xhosa, Yiddish.

Thus, results for language pairs containing uncovered languages are unreliable!

We recommend the two following models to evaluate your translations:

• wmt20-comet-da: DEFAULT Reference-based Regression model build on top of XLM-R (large) and trained of Direct Assessments from WMT17 to WMT19. Same as wmt-large-da-estimator-1719 from previous versions.
• wmt21-comet-qe-mqm: Reference-FREE Regression model build on top of XLM-R (large), trained on Direct Assessments and fine-tuned on MQM.

The default model was developed to participate in the WMT20 Metrics shared task (Mathur et al. 2020) and were among the best metrics that year. Also, in a large-scale study performed by Microsoft Research this metrics ranked 1st in terms of system-level decision accuracy (Kocmi et al. 2020).

Our recommended QE system was developed for the WMT21 Metrics shared task and was the best performing QE as a Metric that year (Freitag et al. 2021).

Note: The range of scores between different models can be totally different. To better understand COMET scores please take a look at our FAQs

For more information about the available COMET models read our metrics descriptions here

Instead of using pretrained models your can train your own model with the following command:

comet-train --cfg configs/models/{your_model_config}.yaml

You can then use your own metric to score:

comet-score -s src.de -t hyp1.en -r ref.en --model PATH/TO/CHECKPOINT

Note: Please contact [email protected] if you wish to host your own metric within COMET available metrics!

In order to run the toolkit tests you must run the following command:

coverage run --source=comet -m unittest discover
coverage report -m

If you use COMET please cite our work! Also, don't forget to say which model you used to evaluate your systems.

• Are References Really Needed? Unbabel-IST 2021 Submission for the Metrics Shared Task

• Uncertainty-Aware Machine Translation Evaluation

• COMET - Deploying a New State-of-the-art MT Evaluation Metric in Production

• Unbabel's Participation in the WMT20 Metrics Shared Task

• COMET: A Neural Framework for MT Evaluation