wmt22-cometkiwi-da-pruned-k2

A compressed version of Unbabel/wmt22-cometkiwi-da — a reference-free machine-translation quality estimation model (source + MT only, no human reference required).

Highest-quality pruned variant — only 1 point human-Pearson loss with 2 encoder layers removed.

What's different from the base model

• 2 encoder layers dropped (indices 4, 5) out of 24. Layer selection by cosine similarity between each layer's input and output hidden states on a small multilingual calibration set.
• layerwise_attention rebuilt to mix only the surviving layers (embeddings + kept layer outputs).
• No quantization — encoder weights remain fp32.

Accuracy

Benchmarked on 1200 stratified segments from RicardoRei/wmt-da-human-evaluation (reference-free, src+mt only):

Metric	This variant	Full cometkiwi
Pearson r vs human DA	0.6300	0.6402
Spearman vs human DA	0.6556	0.6698
Pearson r vs full	0.9784	1.0000
MAE vs full	0.0211	0.0000
Params	539.9M	565.1M
On-disk size	~2160 MB	~2200 MB

All variants at a glance

Variant	Pearson(human)	Pearson(full)	Size	When to use
full base	0.6402	1.0000	~2200 MB	reference quality
-int8	0.6404	0.9919	~1300 MB	lossless compression
-pruned-k2	0.6300	0.9784	~2100 MB	best-quality pruned
-pruned-k4	0.5642	0.8316	~2060 MB	aggressive prune
-pruned-k4-xs	0.5544	0.8113	~1030 MB	smallest footprint

Usage

Standalone — no gated base-model download. The repo ships everything the loader needs (hparams.yaml + state_dict.pt); the loader instantiates an empty COMET architecture via load_pretrained_weights=False and overlays the fine-tuned weights. Only the ungated microsoft/infoxlm-large tokenizer/config (~5 MB) is fetched on first load and cached.

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

from huggingface_hub import snapshot_download
import sys
folder = snapshot_download(repo_id="solailabs/wmt22-cometkiwi-da-pruned-k2")
sys.path.insert(0, folder)
from load import load_model

model = load_model(folder)
out = model.predict(
    [{{"src": "The meeting has been postponed until next week.",
       "mt":  "La réunion a été reportée à la semaine prochaine."}}],
    batch_size=8, gpus=0, progress_bar=False, num_workers=2,
)
print(out["scores"])

No HF_TOKEN required. No license acceptance on Unbabel/wmt22-cometkiwi-da needed.

Files

• state_dict.pt — model weights (fp32 for -pruned-k2 / -pruned-k4, fp16 for -int8 / -pruned-k4-xs)
• hparams.yaml — COMET hyper-parameters (encoder model, regressor shape, loss config)
• config.json — kept/dropped layer indices, quant flag, benchmarked accuracy
• load.py — drop-in standalone loader
• README.md — this file

Citation

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

@inproceedings{{rei-etal-2022-cometkiwi,
    title = "{{C}}omet{{K}}iwi: {{IST}}-{{U}}nbabel 2022 Submission for the Quality Estimation Shared Task",
    author = "Rei, Ricardo  and others",
    booktitle = "WMT 2022",
}}

Released under the same license as the base model (Apache 2.0).