Unbabel COMET

2.2.7 verified Sat May 09 auth: no python

High-quality Machine Translation Evaluation library using neural models. Version 2.2.7. Active development on GitHub, with minor releases every few months.

pip install unbabel-comet

Common errors

error KeyError: 'src' ↓

cause Missing required key in input data dict.

fix

Ensure each sample dict has 'src' and 'mt' keys.

error requests.exceptions.ConnectionError: HTTPConnectionPool ↓

cause Network issue downloading model from Hugging Face hub.

fix

Check internet connection or set HF_HUB_OFFLINE=1 to use cached models.

error AttributeError: module 'comet' has no attribute 'download_model' ↓

cause Incorrect import path.

fix

Use 'from comet import download_model' instead of 'from comet.models import download_model'.

Warnings

gotcha The `predict` method expects a list of dicts with keys 'src', 'mt', and optionally 'ref'. Missing keys cause KeyError. ↓

fix Ensure each dict contains 'src' and 'mt' keys (and 'ref' if required).

gotcha Model downloads can be large (several GB). The download_model function caches in ~/.cache/comet by default. ↓

fix Set environment variable COMET_CACHE_DIR to a custom path.

deprecated The `comet-score` CLI uses model.half() by default starting v2.2.5, which slightly changes scores compared to v2.2.4. ↓

fix Upgrade to v2.2.6+ for consistent half-precision behavior.

gotcha Some models (like XCOMET) are only for inference and not for training in current releases. ↓

fix Check model documentation to confirm training support.

Imports

download_model
wrong
```
from comet.models import download_model
```
correct
```
from comet import download_model
```
download_model is at the package level, not in comet.models

load_from_checkpoint

wrong

from comet.models import load_from_checkpoint

correct

from comet import load_from_checkpoint

Same as download_model; package-level function

Quickstart

Load a COMET model and score translation pairs.

from comet import download_model, load_from_checkpoint

model_path = download_model("Unbabel/wmt22-comet-da")
model = load_from_checkpoint(model_path)

data = [
    {
        "src": "This is a source sentence.",
        "mt": "Ceci est une phrase source.",
        "ref": "This is a reference translation."
    }
]

seg_scores, sys_score = model.predict(data, batch_size=8, gpus=1)
print(seg_scores, sys_score)