This is the converted model from Unbabel/wmt23-cometkiwi-da

Just kept the weights/bias keys()
Renamed the keys to match the original Facebook/XLM-roberta-XL
kept the layer_wise_attention / estimator layers

Because of a hack in HF's code I had to rename the "layerwise_attention.gamma" key to "layerwise_attention.gam"

I changed the config.json key "layer_transformation" from sparsemax to softmax because there is a bug in COMET since the flag is not passed, the actual function used is the default which is softmax.

Usage:

from transformers import XLMRobertaTokenizer, XLMRobertaTokenizerFast, AutoModel
tokenizer = XLMRobertaTokenizerFast.from_pretrained("vince62s/wmt23-cometkiwi-da-roberta-xl", trust_remote_code=True)
model = AutoModel.from_pretrained("vince62s/wmt23-cometkiwi-da-roberta-xl", trust_remote_code=True)

text = "Hello world!</s></s>Bonjour le monde"
encoded_text = tokenizer(text, return_tensors='pt')
print(encoded_text)
output = model(**encoded_text)
print(output[0])

{'input_ids': tensor([[    0, 35378,  8999,    38,     2,     2, 84602,    95, 11146,     2]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])}
tensor([[0.8217]], grad_fn=<AddmmBackward0>)

Let's double check with the original code from Unbabel Comet:

from comet import download_model, load_from_checkpoint
model = load_from_checkpoint("/home/vincent/Downloads/cometkiwi23/checkpoints/model.ckpt") # this is the Unbabel checkpoint
data = [{"mt": "Hello world!", "src": "Bonjour le monde"}]
output = model.predict(data, gpus=0)
print(output)

Prediction([('scores', [0.8216837048530579]), ('system_score', 0.8216837048530579)])

extra_gated_heading: Acknowledge license to accept the repository extra_gated_button_content: Acknowledge license pipeline_tag: translation language: - multilingual - af - am - ar - as - az - be - bg - bn - br - bs - ca - cs - cy - da - de - el - en - eo - es - et - eu - fa - fi - fr - fy - ga - gd - gl - gu - ha - he - hi - hr - hu - hy - id - is - it - ja - jv - ka - kk - km - kn - ko - ku - ky - la - lo - lt - lv - mg - mk - ml - mn - mr - ms - my - ne - nl - 'no' - om - or - pa - pl - ps - pt - ro - ru - sa - sd - si - sk - sl - so - sq - sr - su - sv - sw - ta - te - th - tl - tr - ug - uk - ur - uz - vi - xh - yi - zh license: cc-by-nc-sa-4.0 library_name: transformers

This is a COMET quality estimation model: It receives a source sentence and the respective translation and returns a score that reflects the quality of the translation.

Paper

CometKiwi: IST-Unbabel 2022 Submission for the Quality Estimation Shared Task (Rei et al., WMT 2022)

License:

cc-by-nc-sa-4.0

Usage (unbabel-comet)

Using this model requires unbabel-comet to be installed:

pip install --upgrade pip  # ensures that pip is current 
pip install "unbabel-comet>=2.0.0"

Make sure you acknowledge its License and Log in into Hugging face hub before using:

huggingface-cli login
# or using an environment variable
huggingface-cli login --token $HUGGINGFACE_TOKEN

Then you can use it through comet CLI:

comet-score -s {source-input}.txt -t {translation-output}.txt --model Unbabel/wmt22-cometkiwi-da

Or using Python:

from comet import download_model, load_from_checkpoint

model_path = download_model("Unbabel/wmt22-cometkiwi-da")
model = load_from_checkpoint(model_path)
data = [
    {
        "src": "The output signal provides constant sync so the display never glitches.",
        "mt": "Das Ausgangssignal bietet eine konstante Synchronisation, so dass die Anzeige nie stört."
    },
    {
        "src": "Kroužek ilustrace je určen všem milovníkům umění ve věku od 10 do 15 let.",
        "mt": "Кільце ілюстрації призначене для всіх любителів мистецтва у віці від 10 до 15 років."
    },
    {
        "src": "Mandela then became South Africa's first black president after his African National Congress party won the 1994 election.",
        "mt": "その後、1994年の選挙でアフリカ国民会議派が勝利し、南アフリカ初の黒人大統領となった。"
    }
]
model_output = model.predict(data, batch_size=8, gpus=1)
print (model_output)

Intended uses

Our model is intented to be used for reference-free MT evaluation.

Given a source text and its translation, outputs a single score between 0 and 1 where 1 represents a perfect translation.

Languages Covered:

This model builds on top of InfoXLM 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!