ACADATA: Parallel Dataset of Academic Data for Machine Translation
Paper • 2510.12621 • Published • 1
Releases
To load a specific version use the
revisionparameter:model = AutoModelForCausalLM.from_pretrained( "BSC-LT/salamandraTA-7b-instruct", revision="v1.0" # omit revision for current release )
SalamandraTA Model Card
SalamandraTA-7b-instruct is a translation LLM that has been instruction-tuned from SalamandraTA-7b-base. The base model results from continually pre-training Salamandra-7b on monolingual and parallel data and has not been published, but is reserved for internal use. SalamandraTA-7b-instruct (v2) is proficient in 40 languages (+ 3 varieties) and is mainly trained to perform general translation tasks at the sentence, paragraph, and document levels. The language coverage for this version of SalamandraTA-7B-instruct has been expanded to include five additional non-European languages: Arabic, Japanese, Hindi, Korean, and Simplified Chinese, which complement the European languages supported in our previous model's version and broaden the model's multilingual reach. Additionally, translation performance in all pre-existing language pairs has been improved with respect to the previous model's version (see Evaluation).
DISCLAIMER: This version of Salamandra is tailored exclusively for translation tasks. It lacks chat capabilities and has not been trained with any chat instructions.
Model Details
Description
SalamandraTA-7b-base is a continual pre-training of Salamandra-7b.
The model was trained on a mixture of monolingual and parallel corpora. The continual pre-training was conducted in two stages:
- Stage 1: Monolingual data (≈69B tokens)
- Stage 2: A mixture of monolingual and parallel data, including instruction prompts (≈262B tokens)
In total, approximately 331B tokens were processed during continual pre-training.
Architecture
| Total Parameters | 7,768,117,248 |
| Embedding Parameters | 1,048,576,000 |
| Layers | 32 |
| Hidden size | 4,096 |
| Attention heads | 32 |
| Context length | 8,192 |
| Vocabulary size | 256,000 |
| Precision | bfloat16 |
| Embedding type | RoPE |
| Activation Function | SwiGLU |
| Layer normalization | RMS Norm |
| Flash attention | ✅ |
| Grouped Query Attention | ✅ |
| Num. query groups | 8 |
Intended Use
Direct Use
The model is intended for both research and commercial use in any of the languages included in the training data for general machine translation tasks.
Out-of-scope Use
The model is not intended for malicious activities, such as harming others or violating human rights. Any downstream application must comply with current laws and regulations. Irresponsible usage in production environments without proper risk assessment and mitigation is also discouraged.
Hardware and Software
Training Framework
SalamandraTA-7b-base was continually pre-trained using NVIDIA's NeMo Framework, which leverages PyTorch Lightning for efficient model training in highly distributed settings.
SalamandraTA-7b-instruct was produced with FastChat.
Compute Infrastructure
All models were trained on MareNostrum 5, a pre-exascale EuroHPC supercomputer hosted and operated by Barcelona Supercomputing Center.
The accelerated partition is composed of 1,120 nodes with the following specifications:
- 4x Nvidia Hopper GPUs with 64GB HBM2 memory
- 2x Intel Sapphire Rapids 8460Y+ at 2.3Ghz and 32c each (64 cores)
- 4x NDR200 (BW per node 800Gb/s)
- 512 GB of Main memory (DDR5)
- 460GB on NVMe storage
How to use
The model can be used either directly in Python using the transformers library or deployed as a service and used through standard API calls.
While the former gives the most control over the inference process it requires the code to be executed on a machine with a sufficiently powerful GPU to run the model locally, and is more error prone than the alternative. We therefore strongly recommend the latter, as deploying the model as a service can be done either locally or on a remote server and makes the model available to multiple clients in parallel among other advantages.
Unless you have very specific needs (e.g. for research) that require adapting the inference process it is preferable to follow the "deployment as a service" guidelines below.
Local inference with Python / transformers
You can translate between the following 40 languages (and 3 varieties):
Arabic, Aragonese, Asturian, Basque, Bulgarian, Catalan (and Catalan-Valencian variety), Chinese (simplified), Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, Galician, German, Greek, Hindi, Hungarian, Irish, Italian, Japanese, Korean, Latvian, Lithuanian, Maltese, Norwegian (Bokmål and Nynorsk varieties), Occitan (and Aranese variety), Polish, Portuguese, Romanian, Russian, Serbian, Slovak, Slovenian, Spanish, Swedish, Ukrainian, Welsh.
The instruction-following model uses the commonly adopted ChatML template:
<|im_start|>system
{SYSTEM PROMPT}<|im_end|>
<|im_start|>user
{USER PROMPT}<|im_end|>
<|im_start|>assistant
{MODEL RESPONSE}<|im_end|>
<|im_start|>user
[...]
The easiest way to apply it is by using the tokenizer's built-in functions, as shown in the following snippet.
from datetime import datetime
from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch
model_id = "BSC-LT/salamandraTA-7b-instruct"
source = 'Spanish'
target = 'Catalan'
sentence = "Ayer se fue, tomó sus cosas y se puso a navegar. Una camisa, un pantalón vaquero y una canción, dónde irá, dónde irá. Se despidió, y decidió batirse en duelo con el mar. Y recorrer el mundo en su velero. Y navegar, nai-na-na, navegar"
text = f"Translate the following text from {source} into {target}.\n{source}: {sentence} \n{target}:"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
torch_dtype=torch.bfloat16
)
message = [ { "role": "user", "content": text } ]
date_string = datetime.today().strftime('%Y-%m-%d')
prompt = tokenizer.apply_chat_template(
message,
tokenize=False,
add_generation_prompt=True,
date_string=date_string
)
inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
input_length = inputs.shape[1]
outputs = model.generate(input_ids=inputs.to(model.device),
max_new_tokens=400,
early_stopping=True,
num_beams=5)
print(tokenizer.decode(outputs[0, input_length:], skip_special_tokens=True))
# Ahir se'n va anar, va recollir les seves coses i es va fer a la mar. Una camisa, uns texans i una cançó, on anirà, on anirà. Es va acomiadar i va decidir batre's en duel amb el mar. I fer la volta al món en el seu veler. I navegar, nai-na-na, navegar
Using this template, each turn is preceded by a <|im_start|> delimiter and the role of the entity
(either user, for content supplied by the user, or assistant for LLM responses), and finished with the <|im_end|> token.
General translation
For machine translation tasks, you can use the following prompt template:
Translate the following text from {source} into {target}.
{source}: {source sentence}
{target}:
Show an example
source = 'Catalan'
target = 'Galician'
source_sentence = "Als antics egipcis del període de l'Imperi Nou els fascinaven els monuments dels seus predecessors, que llavors tenien més de mil anys."
text = f"Translate the following text from {source} into {target}.\n{source}: {source_sentence} \n{target}:"
# Os antigos exipcios do período do Imperio Novo estaban fascinados polos monumentos dos seus predecesores, que entón tiñan máis de mil anos de antigüidade.
Deployment as service and remote use (Messages API)
In our experience, vllm works well for deploying the full unquantized version of the model, whereas llama.cpp is appropriate for the quantized (GGUF) version.
We strongly discourage using ollama as we have encountered compatibility issues that may seriously degrade the model's performance.
The easiest and most reliable way to have a working deployment of ALIA-40b-instruct is through the "Deploy / HF Inference Endpoints" option directly on the Hugging Face model page. This automatically creates a functioning endpoint, using vllm or llama.cpp according to the model variant, with an appropriately dimensioned GPU. While there are additional settings available for the endpoint we found the standard configuration proposed by Hugging Face to be a reasonable starting point.
Once the endpoint is running, the model can be easily called using OpenAI's "Messages API" (the de facto standard API for LLM use). By using this API the chat template is applied automatically by the service, requiring no explicit configuration on the client side. The endpoint's configuration page on Hugging Face also provides a "Playground" for testing and API examples, as well as a simple chat interface.
Note that when using the model through the API you need to provide the prompt in the same format as described in the previous section.
Example usage:
# pip install openai
from openai import OpenAI
client = OpenAI(
base_url = YOUR_ENDPOINT_URL,
api_key = YOUR_HF_TOKEN
)
model_id = "BSC-LT/salamandraTA-7b-instruct"
source = 'Spanish'
target = 'Catalan'
sentence = "Ayer se fue, tomó sus cosas y se puso a navegar. Una camisa, un pantalón vaquero y una canción, dónde irá, dónde irá. Se despidió, y decidió batirse en duelo con el mar. Y recorrer el mundo en su velero. Y navegar, nai-na-na, navegar"
text = f"Translate the following text from {source} into {target}.\n{source}: {sentence} \n{target}:"
chat_completion = client.chat.completions.create(
model = model_id,
messages = [
{
"role": "user",
"content": text
}
],
max_tokens = 1000,
temperature = 0.0
)
print(chat_completion.choices[0].message.content)
The model can also be deployed locally or on any server infrastructure with sufficient GPUs, using vllm or llama.cpp. We recommend an initial deployment on Hugging Face as a point of reference and comparison to make sure the model is behaving as expected in the desired deployment setup.
Compatibility wrapper
In order to integrate SalamandraTA as a drop-in replacement for translation services such as Google Translate or DeepL you can use the wrapper provided at https://github.com/langtech-bsc/mt-wrapper . This service accepts incoming requests in Google Translate or DeepL format and translates them to appropriately formatted requests to a SalamandraTA endpoint.
The wrapper service can be deployed locally or on any hosting platform with minimal resource requirements.
Data
Pretraining Data
The model was trained through two stages of continual pre-training using a combination of monolingual and parallel corpora covering all the 40 supported languages and 3 varieties.
In total, approximately 331B tokens were processed during continual pre-training.
Continual Pre-training Stage 1 (CPT1)
The first stage of continual pre-training used only monolingual data.
The corpus is primarily composed of FineWeb2 collections and Wikipedia dumps, covering European languages as well as several additional languages. Total tokens processed in CPT1: ≈69B tokens.
Primary data sources include:
- FineWeb2 multilingual web corpora
- Wikipedia dumps
- Additional internally curated monolingual corpora for low-resource languages such as Aranese.
Continual Pre-training Stage 2 (CPT2)
The second stage of continual pre-training used a mixture of monolingual and parallel data. The total number of tokens processed during CPT2 is 262B tokens.
The training data consists of:
- Monolingual corpora (FineWeb2 and Wikipedia)
- Parallel corpora formatted with instruction-style prompts
The parallel data contains hundreds of language pairs centered around Catalan, Spanish and English, as well as several direct translation pairs between other languages. This highly multilingual corpus is predominantly composed of data sourced from OPUS, with additional data taken from the NTEU Project, Aina Project, and other sources (see: Data Sources and References). Where little parallel Catalan <-> xx data could be found, synthetic Catalan data was generated from the Spanish side of the collected Spanish <-> xx corpora using Projecte Aina's Spanish-Catalan model.
Click the expand button below to see the full list of corpora included in the monolingual training data.
Monolingual Data Sources (CPT1 and CPT2)
| Language | FineWeb2 Web Corpus | Wikipedia (Wikimedia Dumps) | Other Sources |
|---|---|---|---|
| Arabic | ✓ | ||
| Aragonese | ✓ | ||
| Asturian | ✓ | ✓ | |
| Bulgarian | ✓ | ||
| Catalan | ✓ | ||
| Czech | ✓ | ||
| Welsh | ✓ | ✓ | |
| Danish | ✓ | ||
| German | ✓ | ||
| Greek | ✓ | ||
| Spanish | ✓ | ||
| Estonian | ✓ | ||
| Basque | ✓ | ||
| Finnish | ✓ | ||
| French | ✓ | ||
| Irish | ✓ | ✓ | |
| Galician | ✓ | ||
| Hindi | ✓ | ||
| Croatian | ✓ | ||
| Hungarian | ✓ | ||
| Icelandic | ✓ | ||
| Italian | ✓ | ||
| Japanese | ✓ | ||
| Korean | ✓ | ||
| Lithuanian | ✓ | ||
| Latvian | ✓ | ||
| Maltese | ✓ | ✓ | |
| Norwegian Bokmål | ✓ | ||
| Dutch | ✓ | ||
| Norwegian Nynorsk | ✓ | ||
| Occitan | ✓ | ✓ | |
| Polish | ✓ | ||
| Portuguese | ✓ | ||
| Romanian | ✓ | ||
| Russian | ✓ | ||
| Slovak | ✓ | ||
| Slovenian | ✓ | ||
| Serbian | ✓ | ||
| Swedish | ✓ | ||
| Ukrainian | ✓ | ||
| Chinese | ✓ | ||
| English | ✓ | ✓ | |
| Aranese | Catalan-Aranese parallel corpus |
Click the expand button below to see the full list of corpora included in the parallel training data.
Data Sources
Datasets with "-BSC" in their names (e.g., BOUA-SYNTH-BSC, DOGV-SYNTH-BSC) are synthetic datasets obtained by machine translating pre-existing monolingual corpora with our own seq-to-seq models. These datasets were generated internally for model training and are not published.
Catalan pairs (Ca-xx)
| Dataset | Languages |
|---|---|
| AINA | en |
| ARANESE-SYNTH-CORPUS-BSC | arn |
| CCMatrix | eu |
| EUBookshop | lt, pl, pt |
| KDE4 | bg, cs, da, de, el, et, eu, fi, fr, ga, gl, hr, it, ja, lt, lv, nl, pl, pt, ro, sk, sl, sv |
| GlobalVoices | bg, de, fr, it, nl, pl, pt |
| GNOME | eu, fr, ga, gl, pt |
| MaCoCu | en |
| MultiCCAligned | bg, cs, de, el, et, fi, fr, hr, hu, it, ja, lt, lv, nl, pl, ro, sk, sv |
| MultiHPLT | en, et, fi, ga, hr, mt |
| MultiParaCrawl | bg, da |
| NLLB | bg, da, el, en, et, fi, fr, gl, hu, it, lt, lv, pt, ro, sk, sl |
| OpenSubtitles | bg, cs, da, de, el, et, eu, fi, gl, hr, hu, ja, lt, lv, nl, pl, pt, ro, sk, sl, sv |
| OPUS-100 | en |
| Tatoeba | de, pt |
| WikiMatrix | bg, cs, da, de, el, et, eu, fi, fr, gl, hr, hu, it, lt, nl, pl, pt, ro, sk, sl, sv |
| XLENT | eu, ga, gl, ja |
| Catalan-Aranese Parallel Corpus | arn |
Spanish pairs (Es-xx)
| Dataset | Languages |
|---|---|
| BOUA-SYNTH-BSC | val |
| BOUMH | val |
| BOUA-PILAR | val |
| DGT | bg, cs, da, de, el, et, fi, fr, ga, hr, hu, lt, lv, mt, nl, pl, pt, ro, sk, sl, sv |
| DOGV-SYNTH-BSC | val |
| DOGV-PILAR | val |
| EMEA | bg, cs, da, el, fi, hu, lt, lv, mt, nl, pl, ro, sk, sl, sv |
| EUBookshop | cs, da, de, el, fi, fr, ga, it, lv, mt, nl, pl, pt, ro, sk, sl, sv |
| Europarl | bg, cs, da, el, en, fi, fr, hu, lt, lv, nl, pl, pt, ro, sk, sl, sv |
| Europat | en, hr |
| GlobalVoices | bg, de, fr, pt |
| JRC-Acquis | cs, da, et, fr, lt, lv, mt, nl, pl, ro, sv |
| KDE4 | bg, ga, hr |
| LES-CORTS-VALENCIANES-SYNTH-BSC | val |
| MultiCCAligned | bg, fi, fr, hi, hr, it, ja, lv, nl, pt, zh |
| MultiParaCrawl | de, en, fr, ga, hr, hu, it, mt, pt, zh |
| MultiUN | fr, zh |
| News-Commentary | fr, zh |
| NLLB | ar, bg, cs, da, de, el, et, fi, fr, hi, hu, it, ja, lt, lv, nl, pl, pt, ro, sk, sl, sv, zh |
| NTEU | bg, cs, da, de, el, en, et, fi, fr, ga, hr, hu, it, lt, lv, mt, nl, pl, pt, ro, sk, sl, sv |
| OpenSubtitles | ar, da, de, fi, fr, hi, hr, hu, it, ja, lv, nl, zh |
| PILAR-VALENCIAN-AUTH | val |
| PILAR-VALENCIAN-SYNTH | val |
| Tatoeba | pt, zh |
| TildeModel | bg |
| UNPC | ar, en, fr, zh |
| WikiMatrix | bg, en, fr, hr, it, pt, zh |
English pairs (En-xx)
| Dataset | Languages |
|---|---|
| CCMatrix | ga |
| DGT | cs, da, et, ga, hr, hu, lt, lv, mt, sh, sl |
| EMEA | et, hr, lv, mt, ro, sk, sl |
| EUBookshop | ar, cy, ga, is, ja, ru, sh, uk, zh |
| Europarl | cs, et |
| Europat | no |
| GNOME | cy, ga, nn |
| HPLT | hi |
| KDE4 | ar, cy, ga, is, ja, ko, nn, oc, ru, sh, uk, zh |
| GlobalVoices | ar, ja, ko, ru, sh |
| MaCoCu | hr, mt, uk |
| MultiCCAligned | ar, bg, cy, da, et, fi, hr, hu, is, ja, ko, lt, lv, no, ru, sh, sl, sr, uk, zh |
| MultiHPLT | ar, fi, ga, gl, hr, is, ja, ko, mt, nn, sh, sr, uk |
| MultiParaCrawl | bg, cs, da, de, el, et, fi, fr, ga, hr, hu, is, lt, lv, mt, nn, pl, ro, ru, sk, sl, uk |
| MultiUN | ar, ru, zh |
| News-Commentary | ar, cs, ja, ru, zh |
| NLLB | ar, bg, cs, cy, da, de, el, et, fi, fr, ga, hi, hr, hu, it, ja, ko, lt, lv, mt, nl, no, oc, pl, pt, ro, ru, sh, sk, sl, sr, sv, uk, zh |
| NÓS Authentic Corpus | gl |
| NÓS Synthetic Corpus | gl |
| NTEU | da, et, ga, hr, lt, lv, mt, ro, sk, sl, sv |
| Anuvaad | hi |
| OpenSubtitles | ar, bg, cs, de, el, et, fi, fr, hi, hr, hu, is, ja, ko, no, ru, sh, sl, sr, uk, zh |
| OPUS-100 | gl |
| ParaCrawl | cs, et, is, ko, ru, uk, zh |
| ParIce | is |
| Samanantar | hi |
| StanfordNLP-NMT | cs |
| Tatoeba | ar, cs, et, is, ja, ko, ru, sh, uk |
| TildeModel | cs, et, hr, is, lt, lv, mt, ru, sh, uk |
| UNPC | ar, ru, zh |
| Wikimedia | cy, nn |
| XLENT | ar, cs, cy, et, ga, gl, hr, is, ja, ko, oc, ru, sh, uk, zh |
Other pairs
| Dataset | Language pairs |
|---|---|
| DGT | cs-de |
| EUBookshop | cs-de, cs-uk |
| Europarl | cs-de |
| GlobalVoices | cs-de |
| KDE4 | cs-de, cs-uk, ja-zh |
| MultiCCAligned | cs-de, cs-uk, ja-zh |
| MultiParaCrawl | cs-de, cs-uk |
| News-Commentary | cs-de, ja-zh |
| NLLB | cs-de |
| OpenSubtitles | cs-de, cs-uk, ja-zh |
| ParaCrawl | ja-zh |
| Tatoeba | cs-de, cs-uk |
| TildeModel | cs-de |
| WikiMatrix | ja-zh |
| Wikimedia | ja-zh |
| XLENT | cs-de, cs-uk, ja-zh |
To consult the data summary document with the respective licences, please send an e-mail to ipr@bsc.es.
References
- Bañón, M., Chen, P., Haddow, B., Heafield, K., Hoang, H., Esplà-Gomis, M., Forcada, M. L., Kamran, A., Kirefu, F., Koehn, P., Ortiz Rojas, S., Pla Sempere, L., Ramírez-Sánchez, G., Sarrías, E., Strelec, M., Thompson, B., Waites, W., Wiggins, D., & Zaragoza, J. (2020). ParaCrawl: Web-Scale Acquisition of Parallel Corpora. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 4555–4567). Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.417
- de-Dios-Flores, I., Paniagua Suárez, S., Carbajal Pérez, C., Bardanca Outeiriño, D., Garcia, M., & Gamallo, P. (2024). CorpusNÓS: A massive Galician corpus for training large language models. In Proceedings of the 16th International Conference on Computational Processing of Portuguese (Vol. 1, pp. 593–599). https://aclanthology.org/2024.propor-1.66/
- Penedo, G., Kydlíček, H., Sabolčec, V., Messmer, B., Foroutan, N., Kargaran, A. H., Raffel, C., Jaggi, M., Von Werra, L., & Wolf, T. (2025). FineWeb2: One Pipeline to Scale Them All — Adapting Pre-Training Data Processing to Every Language. arXiv. https://arxiv.org/abs/2506.20920
- Wikimedia Foundation. (2025). Wikimedia Downloads. https://dumps.wikimedia.org
- Aulamo, M., Sulubacak, U., Virpioja, S., & Tiedemann, J. (2020). OpusTools and Parallel Corpus Diagnostics. In Proceedings of the Twelfth Language Resources and Evaluation Conference (pp. 3782–3789). European Language Resources Association. https://aclanthology.org/2020.lrec-1.467
- Chaudhary, V., Tang, Y., Guzmán, F., Schwenk, H., & Koehn, P. (2019). Low-Resource Corpus Filtering Using Multilingual Sentence Embeddings. In Proceedings of the Fourth Conference on Machine Translation (Volume 3: Shared Task Papers, Day 2) (pp. 261–266). Association for Computational Linguistics. https://doi.org/10.18653/v1/W19-5435
- DGT-Translation Memory—European Commission. (n.d.). Retrieved November 4, 2024, from https://joint-research-centre.ec.europa.eu/language-technology-resources/dgt-translation-memory_en
- Eisele, A., & Chen, Y. (2010). MultiUN: A Multilingual Corpus from United Nation Documents. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC'10). European Language Resources Association. http://www.lrec-conf.org/proceedings/lrec2010/pdf/686_Paper.pdf
- El-Kishky, A., Chaudhary, V., Guzmán, F., & Koehn, P. (2020). CCAligned: A Massive Collection of Cross-Lingual Web-Document Pairs. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 5960–5969. https://doi.org/10.18653/v1/2020.emnlp-main.480
- El-Kishky, A., Renduchintala, A., Cross, J., Guzmán, F., & Koehn, P. (2021). XLEnt: Mining a Large Cross-lingual Entity Dataset with Lexical-Semantic-Phonetic Word Alignment. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 10424–10430. https://doi.org/10.18653/v1/2021.emnlp-main.814
- Fan, A., Bhosale, S., Schwenk, H., Ma, Z., El-Kishky, A., Goyal, S., Baines, M., Celebi, O., Wenzek, G., Chaudhary, V., Goyal, N., Birch, T., Liptchinsky, V., Edunov, S., Grave, E., Auli, M., & Joulin, A. (2020). Beyond English-Centric Multilingual Machine Translation (No. arXiv:2010.11125). arXiv. https://doi.org/10.48550/arXiv.2010.11125
- Galiano-Jiménez, A., Sánchez-Martínez, F., & Pérez-Ortiz, J. A. (Eds.). (2024). PILAR. GitHub. https://github.com/transducens/PILAR
- García-Martínez, M., Bié, L., Cerdà, A., Estela, A., Herranz, M., Krišlauks, R., Melero, M., O'Dowd, T., O'Gorman, S., Pinnis, M., Stafanovič, A., Superbo, R., & Vasiļevskis, A. (2021). Neural Translation for European Union (NTEU). 316–334. https://aclanthology.org/2021.mtsummit-up.23
- de Gibert, O., Nail, G., Arefyev, N., Bañón, M., van der Linde, J., Ji, S., Zaragoza-Bernabeu, J., Aulamo, M., Ramírez-Sánchez, G., Kutuzov, A., Pyysalo, S., Oepen, S., & Tiedemann, J. (2024). A New Massive Multilingual Dataset for High-Performance Language Technologies (No. arXiv:2403.14009). arXiv. http://arxiv.org/abs/2403.14009
- Koehn, P. (2005). Europarl: A Parallel Corpus for Statistical Machine Translation. Proceedings of Machine Translation Summit X: Papers, 79–86. https://aclanthology.org/2005.mtsummit-papers.11
- Kreutzer, J., Caswell, I., Wang, L., Wahab, A., Van Esch, D., Ulzii-Orshikh, N., Tapo, A., Subramani, N., Sokolov, A., Sikasote, C., Setyawan, M., Sarin, S., Samb, S., Sagot, B., Rivera, C., Rios, A., Papadimitriou, I., Osei, S., Suarez, P. O., … Adeyemi, M. (2022). Quality at a Glance: An Audit of Web-Crawled Multilingual Datasets. Transactions of the Association for Computational Linguistics, 10, 50–72. https://doi.org/10.1162/tacl_a_00447
- Lison, P., & Tiedemann, J. (2016). OpenSubtitles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16) (pp. 923–929). European Language Resources Association. https://aclanthology.org/L16-1147/
- Ramesh, G., Doddapaneni, S., Bheemaraj, A., Jobanputra, M., AK, R., Sharma, A., Sahoo, S., Diddee, H., J, M., Kakwani, D., Kumar, N., Pradeep, A., Nagaraj, S., Deepak, K., Raghavan, V., Kunchukuttan, A., Kumar, P., & Khapra, M. S. (2022). Samanantar: The Largest Publicly Available Parallel Corpora Collection for 11 Indic Languages. Transactions of the Association for Computational Linguistics, 10, 145–162. https://doi.org/10.1162/tacl_a_00452
- Rozis, R., & Skadiņš, R. (2017). Tilde MODEL - Multilingual Open Data for EU Languages. https://aclanthology.org/W17-0235
- Schwenk, H., Chaudhary, V., Sun, S., Gong, H., & Guzmán, F. (2019). WikiMatrix: Mining 135M Parallel Sentences in 1620 Language Pairs from Wikipedia (No. arXiv:1907.05791). arXiv. https://doi.org/10.48550/arXiv.1907.05791
- Schwenk, H., Wenzek, G., Edunov, S., Grave, E., & Joulin, A. (2020). CCMatrix: Mining Billions of High-Quality Parallel Sentences on the WEB (No. arXiv:1911.04944). arXiv. https://doi.org/10.48550/arXiv.1911.04944
- Skadiņš, R., Tiedemann, J., Rozis, R., & Deksne, D. (2014). Billions of Parallel Words for Free: Building and Using the EU Bookshop Corpus. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14) (pp. 1850–1855). European Language Resources Association. https://aclanthology.org/L14-1652/
- Steinberger, R., Pouliquen, B., Widiger, A., Ignat, C., Erjavec, T., Tufiş, D., & Varga, D. (n.d.). The JRC-Acquis: A Multilingual Aligned Parallel Corpus with 20+ Languages. http://www.lrec-conf.org/proceedings/lrec2006/pdf/340_pdf
- Subramani, N., Luccioni, S., Dodge, J., & Mitchell, M. (2023). Detecting Personal Information in Training Corpora: An Analysis. In Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing (TrustNLP 2023) (pp. 208–220). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.trustnlp-1.18
- Tiedemann, J. (2012). Parallel Data, Tools and Interfaces in OPUS. In Proceedings of the Eight International Conference on Language Resources and Evaluation (LREC'12). European Language Resources Association. http://www.lrec-conf.org/proceedings/lrec2012/pdf/463_Paper
- Ziemski, M., Junczys-Dowmunt, M., & Pouliquen, B. (n.d.). The United Nations Parallel Corpus v1.0. https://aclanthology.org/L16-1561
Instruction Tuning Data
This model has been fine-tuned on ~683k instructions, primarily targeting general machine translation tasks. It's important to note that no chat data was used in the fine-tuning process. We created instructions using the following datasets:
- AnCora-Ca-NER
- EIEC (Basque Named Entities Corpus)
- SLI NERC Galician Gold Corpus
- ACAData
- NTEU
- News-Commentary
- NewsPalm
- Tatoeba
- WMT++
- FLEURS
- LangMark
- Q21
- ApeQuest
- Europarl
- Project Gutenberg
- GLITTER
- GeNTE
- EuroGEST
- Eupress (Internally created dataset, in course of publication)
References
References
- Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod, V., Dalmia, S., Riesa, J., Rivera, C., & Bapna, A. (2022). FLEURS: Few-shot learning evaluation of universal representations of speech. 2022 IEEE Spoken Language Technology Workshop (SLT), 798–805. https://arxiv.org/abs/2205.12446
- Deutsch, D., Briakou, E., Caswell, I., Finkelstein, M., Galor, R., Juraska, J., Kovacs, G., Lui, A., Rei, R., Riesa, J., Rijhwani, S., Riley, P., Salesky, E., Trabelsi, F., Winkler, S., Zhang, B., & Freitag, M. (2025). WMT24++: Expanding the language coverage of WMT24 to 55 languages & dialects (No. arXiv:2502.12404). arXiv. https://arxiv.org/abs/2502.12404
- Finkelstein, M., Juraska, J., & Freitag, M. (2024). Introducing the NewsPaLM MBR and QE dataset: LLM-generated high-quality parallel data outperforms traditional web-crawled data (No. arXiv:2408.06537). arXiv. https://arxiv.org/abs/2408.06537
- Ive, J., Specia, L., Szoc, S., Vanallemeersch, T., Van den Bogaert, J., Farah, E., Maroti, C., Ventura, A., & Khalilov, M. (2020). A post-editing dataset in the legal domain: Do we underestimate neural machine translation quality? In N. Calzolari et al. (Eds.), Proceedings of the Twelfth Language Resources and Evaluation Conference (pp. 3692–3697). European Language Resources Association. https://aclanthology.org/2020.lrec-1.455/
- Lacunza, I., Garcia Gilabert, J., De Luca Fornaciari, F., Aula-Blasco, J., Gonzalez-Agirre, A., Melero, M., & Villegas, M. (2025). ACAData: Parallel dataset of academic data for machine translation (No. arXiv:2510.12621). arXiv. https://arxiv.org/abs/2510.12621
- Piergentili, A., Savoldi, B., Fucci, D., Negri, M., & Bentivogli, L. (2023). Hi guys or hi folks? Benchmarking gender-neutral machine translation with the GeNTE corpus. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (pp. 7892–7907). Association for Computational Linguistics. https://arxiv.org/abs/2310.05294
- Project Gutenberg. (n.d.). Project Gutenberg. https://www.gutenberg.org/
- Rowe, J., Klimaszewski, M., Guillou, L., Vallor, S., & Birch, A. (2025). EuroGEST: Investigating gender stereotypes in multilingual language models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. https://arxiv.org/abs/2506.03867
- Specia, L., Harris, K., Blain, F., Burchardt, A., Macketanz, V., Skadiņa, I., Negri, M., & Turchi, M. (2017). Translation quality and productivity: A study on rich morphology languages. Proceedings of Machine Translation Summit XVI, 55–71.
- Tiedemann, J. (2020). The Tatoeba translation challenge – Realistic data sets for low-resource and multilingual MT. Proceedings of the Fifth Conference on Machine Translation, 1174–1182. Association for Computational Linguistics. https://aclanthology.org/2020.wmt-1.139
- Urbizu, G., San Vicente, I., Saralegi, X., Agerri, R., & Soroa, A. (2022). BasqueGLUE: A natural language understanding benchmark for Basque. Proceedings of the Language Resources and Evaluation Conference, 1603–1612. European Language Resources Association. https://aclanthology.org/2022.lrec-1.172
- Velazquez, D., Grace, M., Karageorgos, K., Carin, L., Schliem, A., Zaikis, D., & Wechsler, R. (2025). LangMark: A multilingual dataset for automatic post-editing. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol. 1, pp. 32653–32667). Association for Computational Linguistics. https://aclanthology.org/2025.acl-long.1569/
Evaluation
Below are the evaluation results on the Flores+200 devtest set and BOUQuET test subset (sentence-levl), compared against the state-of-the-art MADLAD400-7B-mt model (Kudugunta, S., et al.), and our previously released translation LLM. These results cover the translation directions CA-XX, ES-XX, EN-XX, as well as XX-CA, XX-ES, and XX-EN. The metrics have been computed excluding Asturian, Aranese, and Aragonese, as we report them separately. We additionally evaluate on the newly added Asian languages (Chinese, Japanese, Korean, Hindi, and Arabic), reported separately with comparisons against MADLAD400-7B-mt model (Kudugunta, S., et al.) and NLLB-200-3.3B (Costa-Jussà, M. R. et al.). The evaluation was conducted using MT-Lens, following the standard setting (beam search with beam size 5, limiting the translation length to 500 tokens). We report the following metrics:
Click to show metrics details
BLEU: Sacrebleu implementation. Signature: nrefs:1— case:mixed— eff:no— tok:13a— smooth:exp—version:2.3.1TER: Sacrebleu implementation.ChrF: Sacrebleu implementation.Comet: Model checkpoint: "Unbabel/wmt22-comet-da".Comet-kiwi: Model checkpoint: "Unbabel/wmt22-cometkiwi-da".Bleurt: Model checkpoint: "lucadiliello/BLEURT-20".MetricX: Model checkpoint: "google/metricx-23-xl-v2p0".MetricX-QE: Model checkpoint: "google/metricx-23-qe-xl-v2p0".
English evaluation
English
This section presents the evaluation metrics for English translation tasks.
Flores+200 devtest set
| Bleu↑ | Ter↓ | ChrF↑ | Comet↑ | Comet-kiwi↑ | Bleurt↑ | MetricX↓ | MetricX-QE↓ | |
|---|---|---|---|---|---|---|---|---|
| EN-XX | ||||||||
| SalamandraTA-7b-instruct (v2) | 37.71 | 49.91 | 64.21 | 0.8919 | 0.8556 | 0.8014 | 0.8491 | 0.7264 |
| SalamandraTA-7b-instruct (v1) | 37.41 | 50.78 | 64.23 | 0.8901 | 0.8561 | 0.7972 | 0.9377 | 0.8045 |
| MADLAD400-7B | 34.69 | 56.78 | 62.14 | 0.8751 | 0.8341 | 0.7790 | 1.5296 | 1.5076 |
| XX-EN | ||||||||
| SalamandraTA-7b-instruct (v2) | 45.87 | 40.98 | 69.24 | 0.8897 | 0.8553 | 0.8036 | 1.0281 | 0.9888 |
| SalamandraTA-7b-instruct (v1) | 44.96 | 41.94 | 68.65 | 0.8884 | 0.8546 | 0.8012 | 1.0419 | 0.9934 |
| MADLAD400-7B | 43.81 | 42.76 | 68.35 | 0.8863 | 0.8554 | 0.7970 | 1.1264 | 1.1430 |
BOUQuET test subset (sentence-levl)
| Bleu↑ | Ter↓ | ChrF↑ | Comet↑ | Comet-kiwi↑ | Bleurt↑ | MetricX↓ | MetricX-QE↓ | |
|---|---|---|---|---|---|---|---|---|
| EN-XX | ||||||||
| SalamandraTA-7b-instruct (v2) | 45.06 | 44.38 | 66.75 | 0.91 | 0.85 | 0.82 | 0.72 | 0.91 |
| SalamandraTA-7b-instruct (v1) | 44.05 | 45.84 | 66.31 | 0.90 | 0.85 | 0.81 | 0.75 | 0.97 |
| MADLAD400-7B | 43.78 | 46.43 | 66.64 | 0.90 | 0.85 | 0.81 | 0.83 | 0.88 |
| XX-EN | ||||||||
| SalamandraTA-7b-instruct (v2) | 49.97 | 38.42 | 68.51 | 0.90 | 0.85 | 0.81 | 0.82 | 0.87 |
| SalamandraTA-7b-instruct (v1) | 48.53 | 39.79 | 67.48 | 0.90 | 0.85 | 0.80 | 0.82 | 0.84 |
| MADLAD400-7B | 49.95 | 38.66 | 68.68 | 0.90 | 0.85 | 0.81 | 0.76 | 0.69 |
Spanish evaluation
Spanish
This section presents the evaluation metrics for Spanish translation tasks.
Flores+200 devtest set
| Bleu↑ | Ter↓ | ChrF↑ | Comet↑ | Comet-kiwi↑ | Bleurt↑ | MetricX↓ | MetricX-QE↓ | |
|---|---|---|---|---|---|---|---|---|
| ES-XX | ||||||||
| SalamandraTA-7b-instruct (v2) | 24.57 | 65.43 | 54.51 | 0.8687 | 0.8212 | 0.7615 | 0.9042 | 0.7890 |
| SalamandraTA-7b-instruct (v1) | 24.04 | 66.92 | 54.31 | 0.8682 | 0.8267 | 0.7599 | 0.9244 | 0.7910 |
| MADLAD400-7B | 21.83 | 72.41 | 52.96 | 0.8578 | 0.8263 | 0.7396 | 1.2570 | 1.3107 |
| XX-ES | ||||||||
| SalamandraTA-7b-instruct (v2) | 25.97 | 60.53 | 53.44 | 0.8529 | 0.8389 | 0.7455 | 0.8861 | 1.1598 |
| SalamandraTA-7b-instruct (v1) | 26.48 | 62.12 | 53.62 | 0.8533 | 0.8378 | 0.7443 | 0.7958 | 1.0700 |
| MADLAD400-7B | 24.90 | 61.75 | 53.04 | 0.8488 | 0.8396 | 0.7390 | 1.0484 | 1.5098 |
BOUQuET test subset (sentence-levl)
| Bleu↑ | Ter↓ | ChrF↑ | Comet↑ | Comet-kiwi↑ | Bleurt↑ | MetricX↓ | MetricX-QE↓ | |
|---|---|---|---|---|---|---|---|---|
| ES-XX | ||||||||
| SalamandraTA-7b-instruct (v2) | 34.67 | 56.05 | 58.58 | 0.89 | 0.82 | 0.79 | 0.74 | 0.85 |
| SalamandraTA-7b-instruct (v1) | 33.15 | 58.43 | 57.74 | 0.89 | 0.82 | 0.79 | 0.75 | 0.85 |
| MADLAD400-7B | 34.60 | 55.99 | 59.07 | 0.88 | 0.82 | 0.79 | 0.79 | 0.82 |
| XX-ES | ||||||||
| SalamandraTA-7b-instruct (v2) | 38.12 | 51.40 | 60.73 | 0.88 | 0.82 | 0.79 | 0.71 | 0.95 |
| SalamandraTA-7b-instruct (v1) | 32.64 | 57.98 | 57.51 | 0.87 | 0.82 | 0.78 | 0.66 | 0.90 |
| MADLAD400-7B | 38.86 | 51.46 | 61.48 | 0.88 | 0.83 | 0.79 | 0.64 | 0.78 |
Catalan evaluation
Catalan
This section presents the evaluation metrics for Catalan translation tasks.
Flores+200 devtest set
| Bleu↑ | Ter↓ | ChrF↑ | Comet↑ | Comet-kiwi↑ | Bleurt↑ | MetricX↓ | MetricX-QE↓ | |
|---|---|---|---|---|---|---|---|---|
| CA-XX | ||||||||
| SalamandraTA-7b-instruct (v2) | 30.32 | 57.69 | 58.67 | 0.8763 | 0.8092 | 0.7777 | 0.9215 | 0.9640 |
| SalamandraTA-7b-instruct (v1) | 30.10 | 58.64 | 58.65 | 0.8760 | 0.8150 | 0.7750 | 0.9631 | 0.9779 |
| MADLAD400-7B | 28.52 | 67.29 | 57.20 | 0.8645 | 0.8111 | 0.7592 | 1.3280 | 1.6205 |
| XX-CA | ||||||||
| SalamandraTA-7b-instruct (v2) | 35.28 | 52.92 | 60.58 | 0.8671 | 0.8129 | 0.7641 | 0.8870 | 1.2876 |
| SalamandraTA-7b-instruct (v1) | 34.72 | 53.96 | 60.25 | 0.8640 | 0.8114 | 0.7577 | 0.9004 | 1.2875 |
| MADLAD400-7B | 33.18 | 54.77 | 59.49 | 0.8605 | 0.8130 | 0.7506 | 1.1917 | 1.7937 |
BOUQuET test subset (sentence-levl)
| Bleu↑ | Ter↓ | ChrF↑ | Comet↑ | Comet-kiwi↑ | Bleurt↑ | MetricX↓ | MetricX-QE↓ | |
|---|---|---|---|---|---|---|---|---|
| CA-XX | ||||||||
| SalamandraTA-7b-instruct (v2) | 31.56 | 61.98 | 56.81 | 0.88 | 0.79 | 0.78 | 0.78 | 1.05 |
| SalamandraTA-7b-instruct (v1) | 31.45 | 60.77 | 56.71 | 0.88 | 0.79 | 0.78 | 0.80 | 1.06 |
| MADLAD400-7B | 32.30 | 58.65 | 56.96 | 0.87 | 0.79 | 0.78 | 0.85 | 1.17 |
| XX-CA | ||||||||
| SalamandraTA-7b-instruct (v2) | 34.50 | 54.62 | 57.69 | 0.86 | 0.77 | 0.76 | 0.90 | 1.43 |
| SalamandraTA-7b-instruct (v1) | 32.99 | 57.49 | 56.26 | 0.86 | 0.78 | 0.76 | 0.91 | 1.39 |
| MADLAD400-7B | 32.19 | 57.27 | 55.85 | 0.85 | 0.77 | 0.75 | 1.00 | 1.56 |
Low-Resource Languages of Spain
The tables below report performance metrics on the FLORES-200 devtest set for translation from English, Spanish, and Catalan into Asturian, Aranese, and Aragonese. Results are compared to Transducens/IbRo-nllb (Galiano Jimenez, et al.).
English evaluation
English-XX
| Model | source | target | Bleu ↑ | Ter ↓ | ChrF ↑ |
|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | en | ast | 34.22 | 51.47 | 62.02 |
| SalamandraTA-7b-instruct (v1) | 31.79 | 54.07 | 61.78 | ||
| Transducens/IbRo-nllb | 20.56 | 63.92 | 53.32 | ||
| SalamandraTA-7b-instruct (v2) | en | arn | 26.20 | 60.90 | 55.21 |
| SalamandraTA-7b-instruct (v1) | 22.77 | 66.06 | 52.61 | ||
| Transducens/IbRo-nllb | 12.81 | 73.21 | 45.76 | ||
| SalamandraTA-7b-instruct (v2) | en | arg | 25.40 | 60.11 | 54.42 |
| SalamandraTA-7b-instruct (v1) | 19.74 | 71.58 | 51.08 | ||
| Transducens/IbRo-nllb | 14.07 | 70.37 | 46.89 |
Spanish evaluation
Spanish-XX
| Model | source | target | Bleu ↑ | Ter ↓ | ChrF ↑ |
|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | es | ast | 23.56 | 65.25 | 54.27 |
| SalamandraTA-7b-instruct (v1) | 20.66 | 71.81 | 53.14 | ||
| Transducens/IbRo-nllb | 16.79 | 76.36 | 50.89 | ||
| SalamandraTA-7b-instruct (v2) | es | arn | 53.33 | 34.71 | 74.43 |
| SalamandraTA-7b-instruct (v1) | 47.37 | 39.29 | 70.65 | ||
| Transducens/IbRo-nllb | 50.20 | 36.60 | 73.16 | ||
| SalamandraTA-7b-instruct (v2) | es | arg | 57.79 | 29.54 | 77.30 |
| SalamandraTA-7b-instruct (v1) | 44.10 | 39.98 | 71.12 | ||
| Transducens/IbRo-nllb | 59.75 | 28.01 | 78.73 |
Catalan evaluation
Catalan-XX
| Model | source | target | Bleu ↑ | Ter ↓ | ChrF ↑ |
|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | ca | ast | 30.41 | 55.92 | 59.29 |
| SalamandraTA-7b-instruct (v1) | 28.13 | 58.84 | 58.98 | ||
| Transducens/IbRo-nllb | 24.77 | 61.60 | 57.49 | ||
| SalamandraTA-7b-instruct (v2) | ca | arn | 31.66 | 55.10 | 59.95 |
| SalamandraTA-7b-instruct (v1) | 30.92 | 54.69 | 59.82 | ||
| Transducens/IbRo-nllb | 31.22 | 54.30 | 60.30 | ||
| SalamandraTA-7b-instruct (v2) | ca | arg | 25.12 | 60.31 | 54.81 |
| SalamandraTA-7b-instruct (v1) | 20.96 | 65.64 | 52.41 | ||
| Transducens/IbRo-nllb | 24.44 | 60.79 | 55.51 |
Chinese, Korean, Japanese, Hindi, Arabic
English evaluation
English → Asian languages
| Model | source | target | BLEU ↑ | ChrF ↑ | COMET ↑ | COMET-kiwi ↑ | BLEURT ↑ | MetricX ↓ | MetricX-QE ↓ |
|---|---|---|---|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | en | zh | 44.32 | 40.08 | 0.88 | 0.84 | 0.73 | 1.64 | 1.56 |
| MADLAD-400-7B | 41.48 | 36.96 | 0.86 | 0.82 | 0.71 | 2.12 | 1.87 | ||
| NLLB-200-3.3B | 27.01 | 25.42 | 0.78 | 0.71 | 0.58 | 5.23 | 7.99 | ||
| SalamandraTA-7b-instruct (v2) | en | ko | 30.18 | 36.91 | 0.89 | 0.86 | 0.68 | 1.82 | 1.33 |
| MADLAD-400-7B | 30.74 | 37.34 | 0.88 | 0.85 | 0.68 | 1.90 | 1.07 | ||
| NLLB-200-3.3B | 28.64 | 35.18 | 0.88 | 0.86 | 0.67 | 1.89 | 0.80 | ||
| SalamandraTA-7b-instruct (v2) | en | ja | 35.23 | 42.11 | 0.91 | 0.87 | 0.71 | 1.37 | 1.12 |
| MADLAD-400-7B | 27.78 | 36.81 | 0.90 | 0.86 | 0.68 | 1.77 | 1.71 | ||
| NLLB-200-3.3B | 20.08 | 30.37 | 0.86 | 0.83 | 0.59 | 2.70 | 2.30 | ||
| SalamandraTA-7b-instruct (v2) | en | hi | 34.06 | 58.17 | 0.81 | 0.85 | 0.73 | 2.00 | 1.23 |
| MADLAD-400-7B | 31.14 | 55.86 | 0.79 | 0.82 | 0.70 | 2.99 | 2.01 | ||
| NLLB-200-3.3B | 33.19 | 58.04 | 0.81 | 0.84 | 0.72 | 2.07 | 1.03 | ||
| SalamandraTA-7b-instruct (v2) | en | ar | 22.89 | 54.69 | 0.86 | 0.81 | 0.72 | 2.06 | 1.70 |
| MADLAD-400-7B | 17.93 | 50.83 | 0.83 | 0.78 | 0.64 | 2.79 | 2.11 | ||
| NLLB-200-3.3B | 19.17 | 51.95 | 0.83 | 0.78 | 0.64 | 2.96 | 1.91 | ||
Asian languages → English
| Model | source | target | BLEU ↑ | ChrF ↑ | COMET ↑ | COMET-kiwi ↑ | BLEURT ↑ | MetricX ↓ | MetricX-QE ↓ |
|---|---|---|---|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | zh | en | 32.98 | 60.68 | 0.88 | 0.84 | 0.77 | 1.27 | 1.33 |
| MADLAD-400-7B | 30.64 | 59.62 | 0.88 | 0.85 | 0.77 | 1.26 | 1.27 | ||
| NLLB-200-3.3B | 29.49 | 57.63 | 0.86 | 0.84 | 0.74 | 1.70 | 1.60 | ||
| SalamandraTA-7b-instruct (v2) | ko | en | 34.08 | 61.29 | 0.89 | 0.85 | 0.77 | 1.33 | 1.35 |
| MADLAD-400-7B | 33.35 | 60.72 | 0.89 | 0.86 | 0.77 | 1.23 | 1.24 | ||
| NLLB-200-3.3B | 29.38 | 57.72 | 0.87 | 0.85 | 0.74 | 1.66 | 1.57 | ||
| SalamandraTA-7b-instruct (v2) | ja | en | 31.50 | 59.87 | 0.88 | 0.86 | 0.76 | 1.38 | 1.20 |
| MADLAD-400-7B | 29.21 | 58.30 | 0.88 | 0.86 | 0.76 | 1.45 | 1.27 | ||
| NLLB-200-3.3B | 28.16 | 56.85 | 0.87 | 0.85 | 0.74 | 1.76 | 1.45 | ||
| SalamandraTA-7b-instruct (v2) | hi | en | 44.95 | 68.29 | 0.90 | 0.85 | 0.78 | 1.31 | 1.42 |
| MADLAD-400-7B | 42.62 | 66.98 | 0.90 | 0.86 | 0.78 | 1.35 | 1.49 | ||
| NLLB-200-3.3B | 42.75 | 67.28 | 0.90 | 0.85 | 0.78 | 1.46 | 1.52 | ||
| SalamandraTA-7b-instruct (v2) | ar | en | 44.89 | 68.32 | 0.88 | 0.81 | 0.78 | 1.37 | 1.66 |
| MADLAD-400-7B | 44.44 | 68.50 | 0.88 | 0.83 | 0.79 | 1.33 | 1.33 | ||
| NLLB-200-3.3B | 42.09 | 66.37 | 0.88 | 0.82 | 0.78 | 1.48 | 1.44 | ||
Spanish evaluation
Spanish → Asian languages
| Model | source | target | BLEU ↑ | ChrF ↑ | COMET ↑ | COMET-kiwi ↑ | BLEURT ↑ | MetricX ↓ | MetricX-QE ↓ |
|---|---|---|---|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | es | zh | 37.02 | 33.75 | 0.87 | 0.80 | 0.71 | 1.77 | 1.58 |
| MADLAD-400-7B | 35.87 | 32.02 | 0.87 | 0.81 | 0.71 | 1.74 | 1.52 | ||
| NLLB-200-3.3B | 20.81 | 20.44 | 0.77 | 0.69 | 0.56 | 5.04 | 8.90 | ||
| SalamandraTA-7b-instruct (v2) | es | ko | 22.42 | 30.27 | 0.87 | 0.82 | 0.66 | 1.81 | 1.32 |
| MADLAD-400-7B | 21.66 | 28.95 | 0.86 | 0.82 | 0.66 | 2.14 | 1.90 | ||
| NLLB-200-3.3B | 20.00 | 27.35 | 0.85 | 0.81 | 0.65 | 2.12 | 0.96 | ||
| SalamandraTA-7b-instruct (v2) | es | ja | 27.63 | 35.35 | 0.89 | 0.83 | 0.68 | 1.57 | 1.55 |
| MADLAD-400-7B | 18.20 | 28.64 | 0.85 | 0.79 | 0.65 | 3.02 | 4.05 | ||
| NLLB-200-3.3B | 15.65 | 25.43 | 0.85 | 0.79 | 0.57 | 2.76 | 2.02 | ||
| SalamandraTA-7b-instruct (v2) | es | hi | 21.22 | 47.37 | 0.76 | 0.79 | 0.68 | 2.24 | 1.40 |
| MADLAD-400-7B | 19.42 | 46.48 | 0.75 | 0.80 | 0.68 | 2.28 | 1.63 | ||
| NLLB-200-3.3B | 18.90 | 45.76 | 0.75 | 0.80 | 0.67 | 2.44 | 1.13 | ||
| SalamandraTA-7b-instruct (v2) | es | ar | 16.34 | 48.55 | 0.85 | 0.79 | 0.69 | 1.71 | 1.39 |
| MADLAD-400-7B | 5.22 | 37.46 | 0.75 | 0.70 | 0.54 | 5.06 | 5.42 | ||
| NLLB-200-3.3B | 8.46 | 40.74 | 0.78 | 0.73 | 0.55 | 3.67 | 2.37 | ||
Asian languages → Spanish
| Model | source | target | BLEU ↑ | ChrF ↑ | COMET ↑ | COMET-kiwi ↑ | BLEURT ↑ | MetricX ↓ | MetricX-QE ↓ |
|---|---|---|---|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | zh | es | 21.59 | 49.80 | 0.85 | 0.82 | 0.71 | 1.15 | 1.62 |
| MADLAD-400-7B | 21.27 | 50.00 | 0.86 | 0.84 | 0.72 | 1.00 | 1.49 | ||
| NLLB-200-3.3B | 18.04 | 46.61 | 0.83 | 0.80 | 0.67 | 1.92 | 2.18 | ||
| SalamandraTA-7b-instruct (v2) | ko | es | 20.92 | 49.41 | 0.85 | 0.84 | 0.71 | 1.17 | 1.59 |
| MADLAD-400-7B | 20.58 | 48.69 | 0.85 | 0.84 | 0.71 | 1.25 | 1.82 | ||
| NLLB-200-3.3B | 17.32 | 45.53 | 0.83 | 0.82 | 0.67 | 1.91 | 2.08 | ||
| SalamandraTA-7b-instruct (v2) | ja | es | 21.00 | 49.09 | 0.85 | 0.85 | 0.71 | 1.20 | 1.46 |
| MADLAD-400-7B | 20.05 | 48.13 | 0.85 | 0.84 | 0.70 | 1.32 | 1.55 | ||
| NLLB-200-3.3B | 17.39 | 45.76 | 0.83 | 0.83 | 0.67 | 1.96 | 1.98 | ||
| SalamandraTA-7b-instruct (v2) | hi | es | 24.46 | 51.68 | 0.86 | 0.82 | 0.72 | 1.10 | 1.73 |
| MADLAD-400-7B | 22.57 | 50.49 | 0.86 | 0.82 | 0.72 | 1.40 | 2.20 | ||
| NLLB-200-3.3B | 20.95 | 48.87 | 0.84 | 0.82 | 0.69 | 1.73 | 2.10 | ||
| SalamandraTA-7b-instruct (v2) | ar | es | 24.59 | 51.75 | 0.84 | 0.80 | 0.72 | 1.33 | 2.08 |
| MADLAD-400-7B | 24.22 | 52.44 | 0.85 | 0.82 | 0.73 | 1.22 | 1.64 | ||
| NLLB-200-3.3B | 21.30 | 50.02 | 0.83 | 0.80 | 0.70 | 1.69 | 1.92 | ||
Catalan evaluation
Catalan → Asian languages
| Model | source | target | BLEU ↑ | ChrF ↑ | COMET ↑ | COMET-kiwi ↑ | BLEURT ↑ | MetricX ↓ | MetricX-QE ↓ |
|---|---|---|---|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | ca | zh | 38.20 | 35.25 | 0.87 | 0.80 | 0.71 | 1.90 | 2.19 |
| MADLAD-400-7B | 39.58 | 35.14 | 0.87 | 0.81 | 0.71 | 1.75 | 1.92 | ||
| NLLB-200-3.3B | 25.42 | 23.57 | 0.80 | 0.72 | 0.60 | 4.06 | 6.20 | ||
| SalamandraTA-7b-instruct (v2) | ca | ko | 26.08 | 33.42 | 0.87 | 0.82 | 0.67 | 1.80 | 1.27 |
| MADLAD-400-7B | 27.15 | 33.15 | 0.88 | 0.83 | 0.69 | 1.62 | 1.14 | ||
| NLLB-200-3.3B | 23.06 | 29.71 | 0.85 | 0.81 | 0.65 | 2.24 | 0.99 | ||
| SalamandraTA-7b-instruct (v2) | ca | ja | 29.20 | 37.48 | 0.90 | 0.82 | 0.69 | 1.43 | 1.52 |
| MADLAD-400-7B | 23.88 | 31.56 | 0.86 | 0.79 | 0.65 | 2.78 | 3.62 | ||
| NLLB-200-3.3B | 16.58 | 26.62 | 0.85 | 0.78 | 0.57 | 2.80 | 2.31 | ||
| SalamandraTA-7b-instruct (v2) | ca | hi | 26.79 | 51.72 | 0.77 | 0.81 | 0.70 | 2.23 | 1.67 |
| MADLAD-400-7B | 23.76 | 50.48 | 0.77 | 0.81 | 0.69 | 2.25 | 1.69 | ||
| NLLB-200-3.3B | 24.11 | 49.92 | 0.75 | 0.80 | 0.68 | 2.58 | 1.58 | ||
| SalamandraTA-7b-instruct (v2) | ca | ar | 18.90 | 50.90 | 0.85 | 0.79 | 0.71 | 1.76 | 1.95 |
| MADLAD-400-7B | 6.03 | 37.68 | 0.74 | 0.67 | 0.53 | 5.98 | 7.53 | ||
| NLLB-200-3.3B | 11.50 | 43.77 | 0.79 | 0.72 | 0.57 | 3.81 | 3.29 | ||
Asian languages → Catalan
| Model | source | target | BLEU ↑ | ChrF ↑ | COMET ↑ | COMET-kiwi ↑ | BLEURT ↑ | MetricX ↓ | MetricX-QE ↓ |
|---|---|---|---|---|---|---|---|---|---|
| SalamandraTA-7b-instruct (v2) | zh | ca | 28.35 | 55.25 | 0.86 | 0.80 | 0.72 | 1.41 | 1.68 |
| MADLAD-400-7B | 26.45 | 54.19 | 0.86 | 0.82 | 0.72 | 1.47 | 1.58 | ||
| NLLB-200-3.3B | 20.85 | 48.92 | 0.83 | 0.78 | 0.64 | 2.73 | 2.66 | ||
| SalamandraTA-7b-instruct (v2) | ko | ca | 27.29 | 54.80 | 0.86 | 0.81 | 0.72 | 1.32 | 2.03 |
| MADLAD-400-7B | 26.01 | 54.07 | 0.86 | 0.82 | 0.72 | 1.30 | 1.95 | ||
| NLLB-200-3.3B | 21.37 | 49.03 | 0.83 | 0.79 | 0.65 | 2.45 | 2.74 | ||
| SalamandraTA-7b-instruct (v2) | ja | ca | 26.30 | 54.21 | 0.86 | 0.81 | 0.71 | 1.50 | 1.87 |
| MADLAD-400-7B | 23.96 | 51.73 | 0.85 | 0.82 | 0.70 | 1.54 | 1.85 | ||
| NLLB-200-3.3B | 20.64 | 48.63 | 0.83 | 0.80 | 0.65 | 2.47 | 2.42 | ||
| SalamandraTA-7b-instruct (v2) | hi | ca | 31.76 | 57.56 | 0.87 | 0.81 | 0.73 | 1.40 | 2.18 |
| MADLAD-400-7B | 31.63 | 57.69 | 0.87 | 0.82 | 0.73 | 1.38 | 1.78 | ||
| NLLB-200-3.3B | 26.33 | 53.33 | 0.85 | 0.81 | 0.68 | 2.22 | 2.47 | ||
| SalamandraTA-7b-instruct (v2) | ar | ca | 33.88 | 59.29 | 0.86 | 0.78 | 0.74 | 1.45 | 1.98 |
| MADLAD-400-7B | 32.92 | 59.28 | 0.86 | 0.80 | 0.74 | 1.53 | 1.70 | ||
| NLLB-200-3.3B | 27.69 | 55.66 | 0.84 | 0.78 | 0.69 | 2.25 | 2.02 | ||
Ethical Considerations and Limitations
Detailed information on the work done to examine the presence of unwanted social and cognitive biases in the base model can be found at Salamandra-7B model card. With regard to MT models, no specific analysis has yet been carried out in order to evaluate potential biases or limitations in translation accuracy across different languages, dialects, or domains. However, we recognize the importance of identifying and addressing any harmful stereotypes, cultural inaccuracies, or systematic performance discrepancies that may arise in Machine Translation. As such, we plan to continue performing more analyses as we implement the necessary metrics and methods within our evaluation framework MT-Lens. Note that the model has only undergone preliminary instruction tuning. We urge developers to consider potential limitations and conduct safety testing and tuning tailored to their specific applications.
Additional information
Author
The Language Technologies Unit from Barcelona Supercomputing Center.
Contact
For further information, please send an email to langtech@bsc.es.
Copyright
Copyright(c) 2025 by Language Technologies Unit, Barcelona Supercomputing Center.
Funding
This work is funded by the Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación, Transformación y Resiliencia - Funded by EU – NextGenerationEU within the framework of the project Desarrollo Modelos ALIA.
This work has been promoted and financed by the Government of Catalonia through the Aina Project.
Acknowledgements
Disclaimer
Be aware that the model may contain biases or other unintended distortions. When third parties deploy systems or provide services based on this model, or use the model themselves, they bear the responsibility for mitigating any associated risks and ensuring compliance with applicable regulations, including those governing the use of Artificial Intelligence.
The Barcelona Supercomputing Center, as the owner and creator of the model, shall not be held liable for any outcomes resulting from third-party use.
License
Citation
- Downloads last month
- 5,006
Model tree for BSC-LT/salamandraTA-7b-instruct
Spaces using BSC-LT/salamandraTA-7b-instruct 3
Papers for BSC-LT/salamandraTA-7b-instruct
FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language
Paper • 2506.20920 • Published • 78
EuroGEST: Investigating gender stereotypes in multilingual language models
Paper • 2506.03867 • Published
WMT24++: Expanding the Language Coverage of WMT24 to 55 Languages & Dialects
Paper • 2502.12404 • Published • 4
Introducing the NewsPaLM MBR and QE Dataset: LLM-Generated High-Quality Parallel Data Outperforms Traditional Web-Crawled Data
Paper • 2408.06537 • Published