COMET-poly-ic1-wmt25

This model is based on COMET-poly, which is a fork but not compatible with original Unbabel's COMET. To run the model, you need to first install this version of COMET either with:

pip install "git+https://github.com/zouharvi/COMET-poly#egg=comet-poly&subdirectory=comet_poly"

or in editable mode:

git clone https://github.com/zouharvi/COMET-poly.git
cd COMET-poly
pip3 install -e comet_poly

This model scores the translation mt but takes additional in-context example: source src2, translation mt2, and score score2, which makes it a better quality estimator:

import comet_poly
model = comet_poly.load_from_checkpoint(comet_poly.download_model("zouharvi/COMET-poly-ic1-wmt25"))
data = [
    {
        "src": "Iceberg lettuce got its name in the 1920s when it was shipped packed in ice to stay fresh.",
        "mt": "Eisbergsalat erhielt seinen Namen in den 1920er-Jahren, als er in Eis verpackt verschickt wurde, um frisch zu bleiben.",
        "src2": "Lettuce is mostly water, which helps keep it crisp when chilled.",
        "mt2": "Kopfsalat besteht größtenteils aus Wasser, was ihm hilft, beim Kühlen knackig zu bleiben.",
        "score2": 94.5
    },
    {
        "src": "Goats have rectangular pupils, which give them a wide field of vision—up to 320 degrees!",
        "mt": "Kozy mají obdélníkové zornice, což jim umožňuje vidět skoro všude kolem sebe, aniž by musely otáčet hlavou.",
        "src2": "Sheep, like goats, also have rectangular pupils for better peripheral vision.",
        "mt2": "Вівці, як і кози, також мають прямокутні зіниці для кращого периферичного зору.",
        "score2": 96.0
    },
    {
        "src": "This helps them spot predators from almost all directions without moving their heads.",
        "mt": "Điều này giúp chúng phát hiện kẻ săn mồi từ gần như mọi hướng mà không cần quay đầu.",
        "src2": "Many prey animals have evolved to detect threats with minimal movement.",
        "mt2": "Nhiều động vật thịt có tiến hóa để xem mối nguy bằng nhỏ đi lại.",
        "score2": 42.3
    }
]
print("scores", model.predict(data, batch_size=8, gpus=1).scores)

Outputs:

scores [98.09857940673828, 85.52458953857422, 83.38972473144531]

You can use a readily-available training data to do the on-the-fly retrieval. Specifically, this model has been trained with retrieval based on src:

import datasets
import comet_poly.retrieval
data = [
    {
        "src": "Iceberg lettuce got its name in the 1920s when it was shipped packed in ice to stay fresh.",
        "mt": "Eisbergsalat erhielt seinen Namen in den 1920er-Jahren, als er in Eis verpackt verschickt wurde, um frisch zu bleiben.",
    },
    {
        "src": "Goats have rectangular pupils, which give them a wide field of vision—up to 320 degrees!",
        "mt": "Kozy mají obdélníkové zornice, což jim umožňuje vidět skoro všude kolem sebe, aniž by musely otáčet hlavou.",
    },
    {
        "src": "This helps them spot predators from almost all directions without moving their heads.",
        "mt": "Điều này giúp chúng phát hiện kẻ săn mồi từ gần như mọi hướng mà không cần quay đầu.",
    }
]

data_kb = list(datasets.load_dataset("zouharvi/wmt-human-all", split="train"))
data_retrieved = comet_poly.retrieval.retrieve_from_kb(
    data=data,
    data_kb=data_kb,
    k=1, # this model takes one in-context example
    prevent_hardmatch=False,
    key="src",
)
# add the retrieved data
for line, lines_retrieved in zip(data, data_retrieved):
    for i in range(len(lines_retrieved)):
        line[f"src{i+2}"] = lines_retrieved[i]["src"]
        line[f"mt{i+2}"] = lines_retrieved[i]["mt"]
        line[f"score{i+2}"] = lines_retrieved[i]["score"]

print("scores", model.predict(data, batch_size=8, gpus=1).scores)

The training data is WMT up to 2024 (inclusive) with DA/ESA/MQM merged on a single scale. This model is based on the work COMET-poly: Machine Translation Metric Grounded in Other Candidates which can be cited as:

@inproceedings{zufle-etal-2025-comet,
    title = "{COMET}-poly: Machine Translation Metric Grounded in Other Candidates",
    author = {Z{\"u}fle, Maike  and
      Zouhar, Vil{\'e}m  and
      Dinh, Tu Anh  and
      Maia Polo, Felipe  and
      Niehues, Jan  and
      Sachan, Mrinmaya},
    editor = "Haddow, Barry  and
      Kocmi, Tom  and
      Koehn, Philipp  and
      Monz, Christof",
    booktitle = "Proceedings of the Tenth Conference on Machine Translation",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.wmt-1.63/",
    doi = "10.18653/v1/2025.wmt-1.63",
    pages = "887--904",
    ISBN = "979-8-89176-341-8",
    abstract = "Automated metrics for machine translation attempt to replicate human judgment. Unlike humans, who often assess a translation in the context of multiple alternatives, these metrics typically consider only the source sentence and a single translation. This discrepancy in the evaluation setup may negatively impact the performance of automated metrics. We propose two automated metrics that incorporate additional information beyond the single translation. COMET-polycand uses alternative translations of the same source sentence to compare and contrast with the translation at hand, thereby providing a more informed assessment of its quality. COMET-polyic, inspired by retrieval-based in-context learning, takes in translations of similar source texts along with their human-labeled quality scores to guide the evaluation. We find that including a single additional translation in COMET-polycand improves the segment-level metric performance (0.079 to 0.118 Kendall{'}s tau-b correlation), with further gains when more translations are added. Incorporating retrieved examples in COMET-polyic yields similar improvements (0.079 to 0.116 Kendall{'}s tau-b correlation). We release our models publicly."
}