wmt22-comet-da-pruned-k4-refit

wmt22-comet-da-pruned-k4-refit

A compressed version of Unbabel/wmt22-comet-da produced by an internal pipeline that removes redundant encoder capacity and re-aligns the layer-mixing weights afterwards. Same disk footprint as the -xs variant, higher quality.

Real-world benchmark

Evaluated on 1,200 WMT17 DA segments across 12 language pairs (RicardoRei/wmt-da-human-evaluation). This is the standard way to judge COMET variants: how well do the model's scores correlate with human quality judgments?

The refit variant recovers roughly 40 % of the quality loss incurred by -xs at the same disk size.

Usage

pip install "unbabel-comet" "setuptools<81" huggingface_hub

from huggingface_hub import snapshot_download
import sys

folder = snapshot_download(repo_id="solailabs/wmt22-comet-da-pruned-k4-refit")
sys.path.insert(0, folder)
from load import load_model

model = load_model()
scores = model.predict(
    [{"src": "Hello world.", "mt": "Bonjour le monde.", "ref": "Bonjour le monde."}],
    batch_size=8, gpus=0, num_workers=2,
)
print(scores["scores"])

The bundled load.py downloads the base model on first use, applies internal compression, and returns a working COMET model. First call is slow (base download ~2.2 GB cached); subsequent calls are fast.

Notes

• No fine-tuning was performed. Only post-hoc structural changes + a short (CPU-only) calibration on a small multilingual sample.
• Tested on Apple M-series (qnnpack quant engine) and x86 Linux (fbgemm).
• Behavior outside the languages listed above is not guaranteed.

License & attribution

Apache-2.0 (inherited from the base model).

Base model: Unbabel/wmt22-comet-da by Unbabel.

@inproceedings{rei-etal-2022-comet,
  title={{COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task}},
  author={Rei, Ricardo and others},
  booktitle={Proceedings of the Seventh Conference on Machine Translation},
  year={2022}
}