Salamandra is a highly multilingual model pre-trained from scratch that comes in three different
sizes — 2B, 7B and 40B parameters — with their respective base and instruction-tuned variants.
This model card corresponds to the 7B instructed version.
To visit the model cards of other Salamandra versions, please refer to the Model Index.
The entire Salamandra family is released under a permissive Apache 2.0 license.
Along with the open weights, all training scripts and configuration files are made publicly available in this GitHub repository.
[!WARNING]
DISCLAIMER: This model is a first proof-of-concept designed to demonstrate the instruction-following capabilities of recently released base models.
It has been optimized to engage in conversation but has NOT been aligned through RLHF to filter or avoid sensitive topics.
As a result, it may generate harmful or inappropriate content.
The team is actively working to enhance its performance through further instruction and alignment with RL techniques.
Model Details
Description
Transformer-based decoder-only language model that has been pre-trained from scratch on 7.8 trillion tokens of highly curated data.
The pre-training corpus contains text in 35 European languages and code.
Hyperparameters
The full list of hyperparameters for each model can be found here.
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 models are intended for both research and commercial use in any of the languages included in the training data.
The base models are intended either for language generation or to be further fine-tuned for specific use-cases.
The instruction-tuned variants can be used as general-purpose assistants, as long as the user is fully aware of the model’s limitations.
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
Pre-training was conducted using NVIDIA’s NeMo Framework,
which leverages PyTorch Lightning for efficient model training in highly distributed settings.
The instruction-tuned versions were 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 64 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
Model
Nodes
GPUs
2B
64
256
7B
128
512
40B
256 / 512
1,024 / 2,048
How to use
The instruction-following models use the commonly adopted ChatML template:
{%- if not date_string is defined %}{%- set date_string = "2024-09-30" %}{%- endif %}{{ "<|im_start|>system\nsystem_message\nToday Date: "+ date_string +"<|im_end|>\n" }}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}
Where system_message is used to guide the model during generation and date_string can be set to allow the model to respond with the current date.
The exact same chat template should be used for an enhanced conversational experience.
The easiest way to apply it is by using the tokenizer's built-in functions, as shown in the following snippet.
python
1from datetime import datetime
2from transformers import AutoTokenizer, AutoModelForCausalLM
3import transformers
4import torch
56model_id ="BSC-LT/salamandra-7b-instruct"78text ="At what temperature does water boil?"910tokenizer = AutoTokenizer.from_pretrained(model_id)11model = AutoModelForCausalLM.from_pretrained(12 model_id,13 device_map="auto",14 torch_dtype=torch.bfloat16
15)1617message =[{"role":"user","content": text }]18date_string = datetime.today().strftime('%Y-%m-%d')1920prompt = tokenizer.apply_chat_template(21 message,22 tokenize=False,23 add_generation_prompt=True,24 date_string=date_string
25)2627inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")28outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=200)2930print(tokenizer.decode(outputs[0], skip_special_tokens=True))
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.
Data
Pretraining Data
The training corpus consists of 2.4 trillion tokens, including 35 European languages and 92 programming languages. It amounts to a total of 33TB of pre-processed text.
Languages were sampled manually by giving x2 oversampling to Spain's co-official languages (Spanish, Catalan, Galician and Basque), code was undersampled by half,
and the rest of the languages were kept as is, resulting in the following distribution:
lang distrib
This highly multilingual corpus is predominantly composed of data from Colossal OSCAR,
which contributes a significant 66.06% of the total tokens.
Following this, Starcoder provides 11.91%, and Spanish Crawling adds 3.34%.
The next largest sources are French FR at 3.12% and Proof Pile at 1.98%.
Other notable contributions include Macocu, Pile of Law, and Eurlex, each contributing around 1.5% to 1.3%.
These major sources collectively form the bulk of the corpus, ensuring a rich and diverse dataset for training the language model.
The remaining 10% comes from smaller sources in various languages.
Feel free to click the expand button below to see the full list of sources.
Data Sources
Dataset
Language
Source
Parlamint corpus
at, bg, cz, dk, ee, es, es-ga, fi, fr, gb, gr, hr, hu, it, lv, nl, no, pl, pt, rs, se, si
Abadji, J., Suárez, P. J. O., Romary, L., & Sagot, B. (2021). Ungoliant: An optimized pipeline for the generation of a very large-scale multilingual web corpus (H. Lüngen, M. Kupietz, P. Bański, A. Barbaresi, S. Clematide, & I. Pisetta, Eds.; pp. 1–9). Leibniz-Institut für Deutsche Sprache. Link
Artetxe, M., Aldabe, I., Agerri, R., Perez-de-Viñaspre, O., & Soroa, A. (2022). Does Corpus Quality Really Matter for Low-Resource Languages?
Bañón, M., Esplà-Gomis, M., Forcada, M. L., García-Romero, C., Kuzman, T., Ljubešić, N., van Noord, R., Sempere, L. P., Ramírez-Sánchez, G., Rupnik, P., Suchomel, V., Toral, A., van der Werff, T., & Zaragoza, J. (2022). MaCoCu: Massive collection and curation of monolingual and bilingual data: Focus on under-resourced languages. Proceedings of the 23rd Annual Conference of the European Association for Machine Translation, 303–304. Link
Brack, M., Ostendorff, M., Suarez, P. O., Saiz, J. J., Castilla, I. L., Palomar-Giner, J., Shvets, A., Schramowski, P., Rehm, G., Villegas, M., & Kersting, K. (2024). Community OSCAR: A Community Effort for Multilingual Web Data. Link
Computer, T. (2023). RedPajama: An Open Source Recipe to Reproduce LLaMA training dataset [Computer software]. Link
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 (arXiv:2403.14009). arXiv. Link
Dodge, J., Sap, M., Marasović, A., Agnew, W., Ilharco, G., Groeneveld, D., Mitchell, M., & Gardner, M. (2021). Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus. In M.-F. Moens, X. Huang, L. Specia, & S. W. Yih (Eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 1286–1305). Association for Computational Linguistics. Link
Erjavec, T., Ljubešić, N., & Logar, N. (2015). The slWaC corpus of the Slovene web. Informatica (Slovenia), 39, 35–42.
Erjavec, T., Ogrodniczuk, M., Osenova, P., Ljubešić, N., Simov, K., Grigorova, V., Rudolf, M., Pančur, A., Kopp, M., Barkarson, S., Steingrímsson, S. hór, van der Pol, H., Depoorter, G., de Does, J., Jongejan, B., Haltrup Hansen, D., Navarretta, C., Calzada Pérez, M., de Macedo, L. D., … Rayson, P. (2021). Linguistically annotated multilingual comparable corpora of parliamentary debates ParlaMint.ana 2.1. Link
Etxaniz, J., Sainz, O., Perez, N., Aldabe, I., Rigau, G., Agirre, E., Ormazabal, A., Artetxe, M., & Soroa, A. (2024). Latxa: An Open Language Model and Evaluation Suite for Basque. [Link] (https://arxiv.org/abs/2403.20266)
Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., Presser, S., & Leahy, C. (2021). The Pile: An 800GB Dataset of Diverse Text for Language Modeling. CoRR, abs/2101.00027. Link
Gutiérrez-Fandiño, A., Armengol-Estapé, J., Gonzalez-Agirre, A., & Villegas, M. (2021). Spanish Legalese Language Model and Corpora.
Hansen, D. H. (2018). The Danish Parliament Corpus 2009—2017, v1. Link
Henderson*, P., Krass*, M. S., Zheng, L., Guha, N., Manning, C. D., Jurafsky, D., & Ho, D. E. (2022). Pile of Law: Learning Responsible Data Filtering from the Law and a 256GB Open-Source Legal Dataset. arXiv. Link
Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., & Steinhardt, J. (2021). Measuring Mathematical Problem Solving With the MATH Dataset. NeurIPS.
Jansen, T., Tong, Y., Zevallos, V., & Suarez, P. O. (2022). Perplexed by Quality: A Perplexity-based Method for Adult and Harmful Content Detection in Multilingual Heterogeneous Web Data.
Koppel, K., & Kallas, J. (2022). Eesti keele ühendkorpuste sari 2013–2021: Mahukaim eestikeelsete digitekstide kogu. Eesti Rakenduslingvistika Ühingu Aastaraamat Estonian Papers in Applied Linguistics, 18, 207–228. Link
Křen, M., Cvrček, V., Henyš, J., Hnátková, M., Jelínek, T., Kocek, J., Kováříková, D., Křivan, J., Milička, J., Petkevič, V., Procházka, P., Skoumalová, H., Šindlerová, J., & Škrabal, M. (2021). SYN v9: Large corpus of written Czech. Link
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. Link
Kummervold, P. E., De la Rosa, J., Wetjen, F., & Brygfjeld, S. A. (2021). Operationalizing a National Digital Library: The Case for a Norwegian Transformer Model. In S. Dobnik & L. Øvrelid (Eds.), Proceedings of the 23rd Nordic Conference on Computational Linguistics (NoDaLiDa) (pp. 20–29). Linköping University Electronic Press, Sweden. Link
Lewandowska-Tomaszczyk, B., Górski, R., Łaziński, M., & Przepiórkowski, A. (2013). The National Corpus of Polish (NKJP). Language use and data analysis. 309–319.
Li, R., Allal, L. B., Zi, Y., Muennighoff, N., Kocetkov, D., Mou, C., Marone, M., Akiki, C., Li, J., Chim, J., Liu, Q., Zheltonozhskii, E., Zhuo, T. Y., Wang, T., Dehaene, O., Davaadorj, M., Lamy-Poirier, J., Monteiro, J., Shliazhko, O., … Vries, H. de. (2023). StarCoder: May the source be with you!
Lison, P., & Tiedemann, J. (2016). OpenSubtitles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In N. Calzolari, K. Choukri, T. Declerck, S. Goggi, M. Grobelnik, B. Maegaard, J. Mariani, H. Mazo, A. Moreno, J. Odijk, & S. Piperidis (Eds.), Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16) (pp. 923–929). European Language Resources Association (ELRA). Link
Ljubešić, N., & Klubička, F. (2014). Bs,hr,srWaC - Web Corpora of Bosnian, Croatian and Serbian. In F. Bildhauer & R. Schäfer (Eds.), Proceedings of the 9th Web as Corpus Workshop (WaC-9) (pp. 29–35). Association for Computational Linguistics. Link
Micallef, K., Gatt, A., Tanti, M., van der Plas, L., & Borg, C. (2022). Pre-training Data Quality and Quantity for a Low-Resource Language: New Corpus and BERT Models for Maltese. Proceedings of the Third Workshop on Deep Learning for Low-Resource Natural Language Processing, 90–101. Link
Ogrodniczuk, M. (2018). Polish Parliamentary Corpus. Link
Ostendorff, M., Blume, T., & Ostendorff, S. (2020). Towards an Open Platform for Legal Information. Proceedings of the ACM/IEEE Joint Conference on Digital Libraries in 2020, 385–388. Link
Ostendorff, M., Suarez, P. O., Lage, L. F., & Rehm, G. (2024). LLM-Datasets: An Open Framework for Pretraining Datasets of Large Language Models. First Conference on Language Modeling. Link
Outsios, S., Skianis, K., Meladianos, P., Xypolopoulos, C., & Vazirgiannis, M. (2018). Word Embeddings from Large-Scale Greek Web content. arXiv Preprint arXiv:1810.06694.
Palomar-Giner, J., Saiz, J. J., Espuña, F., Mina, M., Da Dalt, S., Llop, J., Ostendorff, M., Ortiz Suarez, P., Rehm, G., Gonzalez-Agirre, A., & Villegas, M. (2024). A CURATEd CATalog: Rethinking the Extraction of Pretraining Corpora for Mid-Resourced Languages. In N. Calzolari, M.-Y. Kan, V. Hoste, A. Lenci, S. Sakti, & N. Xue (Eds.), Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 335–349). ELRA and ICCL. Link
Papaloukas, C., Chalkidis, I., Athinaios, K., Pantazi, D.-A., & Koubarakis, M. (2021). Multi-granular Legal Topic Classification on Greek Legislation. Proceedings of the Natural Legal Language Processing Workshop 2021, 63–75. Link
Popa-Fabre, M., Ortiz Suárez, P. J., Sagot, B., & de la Clergerie, É. (2020). French Contextualized Word-Embeddings with a sip of CaBeRnet: A New French Balanced Reference Corpus. Proceedings of the 8th Workshop on Challenges in the Management of Large Corpora, 15–23. Link
Rae, J. W., Potapenko, A., Jayakumar, S. M., Hillier, C., & Lillicrap, T. P. (2019). Compressive Transformers for Long-Range Sequence Modelling. arXiv Preprint. Link
Rodrigues, J., Gomes, L., Silva, J., Branco, A., Santos, R., Cardoso, H. L., & Osório, T. (2023). Advancing Neural Encoding of Portuguese with Transformer Albertina PT-*.
Rødven-Eide, S. (2016). The Swedish Culturomics Gigaword CorpusThe Swedish Culturomics Gigaword Corpus [Dataset]. Språkbanken Text. Link
Sharma, E., Li, C., & Wang, L. (2019). BIGPATENT: A Large-Scale Dataset for Abstractive and Coherent Summarization. CoRR, abs/1906.03741. Link
Soldaini, L., & Lo, K. (2023). peS2o (Pretraining Efficiently on S2ORC) Dataset. Allen Institute for AI.
Strømberg-Derczynski, L., Ciosici, M., Baglini, R., Christiansen, M. H., Dalsgaard, J. A., Fusaroli, R., Henrichsen, P. J., Hvingelby, R., Kirkedal, A., Kjeldsen, A. S., Ladefoged, C., Nielsen, F. Å., Madsen, J., Petersen, M. L., Rystrøm, J. H., & Varab, D. (2021). The Danish Gigaword Corpus. Proceedings of the 23rd Nordic Conference on Computational Linguistics (NoDaLiDa), 413–421. Link
Subramani, N., Luccioni, S., Dodge, J., & Mitchell, M. (2023). Detecting Personal Information in Training Corpora: An Analysis. 208–220. Link
Varab, D., & Schluter, N. (2020). DaNewsroom: A Large-scale Danish Summarisation Dataset. Proceedings of The 12th Language Resources and Evaluation Conference, 6731–6739. Link
Váradi, T., Nyéki, B., Koeva, S., Tadić, M., Štefanec, V., Ogrodniczuk, M., Nitoń, B., Pezik, P., Barbu Mititelu, V., Irimia, E., Mitrofan, M., Tufi\textcommabelows, D., Garabík, R., Krek, S., & Repar, A. (2022). Introducing the CURLICAT Corpora: Seven-language Domain Specific Annotated Corpora from Curated Sources. In N. Calzolari, F. Béchet, P. Blache, K. Choukri, C. Cieri, T. Declerck, S. Goggi, H. Isahara, B. Maegaard, J. Mariani, H. Mazo, J. Odijk, & S. Piperidis (Eds.), Proceedings of the Thirteenth Language Resources and Evaluation Conference (pp. 100–108). European Language Resources Association. Link
Wagner Filho, J. A., Wilkens, R., Idiart, M., & Villavicencio, A. (2018). The brwac corpus: A new open resource for brazilian portuguese. Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018).
Žagar, A., Kavaš, M., Robnik-Šikonja, M., Erjavec, T., Fišer, D., Ljubešić, N., Ferme, M., Borovič, M., Boškovič, B., Ojsteršek, M., & Hrovat, G. (2022). Corpus of academic Slovene KAS 2.0. Link
Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel Bowman. 2022. BBQ: A hand-built bias benchmark for question answering. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2086–2105, Dublin, Ireland. Association for Computational Linguistics.
Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2019. The Woman Worked as a Babysitter: On Biases in Language Generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3407–3412, Hong Kong, China. Association for Computational Linguistics.
Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., & Tafjord, O. (2018). Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge. arXiv:1803. 05457v1.
Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–1642, Seattle, Washington, USA. Association for Computational Linguistics.
Penedo, G., Kydlíček, H., allal, L. B., Lozhkov, A., Mitchell, M., Raffel, C., Von Werra, L., & Wolf, T. (2024). The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale (arXiv:2406.17557). arXiv. http://arxiv.org/abs/2406.17557
Singh, S., Vargus, F., Dsouza, D., Karlsson, B. F., Mahendiran, A., Ko, W.-Y., Shandilya, H., Patel, J., Mataciunas, D., OMahony, L., Zhang, M., Hettiarachchi, R., Wilson, J., Machado, M., Moura, L. S., Krzemiński, D., Fadaei, H., Ergün, I., Okoh, I., … Hooker, S. (2024). Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning (arXiv:2402.06619). arXiv. http://arxiv.org/abs/2402.06619
The model was trained for 3 epochs, with two final rounds of 0.3B higher-quality tokens each,
meaning that the total number of tokens seen during pre-training amounts to roughly 7.8 trillion tokens.
We provide an extense Datasheet section following the best practices defined by (Gebru et al., 2021).
Datasheet
Motivation
For what purpose was the dataset created? Was there a specific task in mind? Was there a specific gap that needed to be filled? Please provide a description.
The purpose of creating this dataset is to pre-train the Salamandra family of multilingual models with high performance in a large number of
European languages (35) and code (including 92 different programming languages). In addition, we aim to represent especially the co-official
languages of Spain: Spanish, Catalan, Galician, and Basque. This is the reason why we carry out an oversampling of these languages.
We detected that there is a great lack of massive multilingual data, especially in minority languages (Ostendorff & Rehm, 2023), so part of
our efforts in the creation of this pre-training dataset have resulted in the contribution to large projects such as the Community OSCAR
(Brack et al., 2024), which includes 151 languages and 40T words, or CATalog (Palomar-Giner et al., 2024), the largest open dataset in
Catalan in the world.
Who created the dataset (e.g., which team, research group) and on behalf of which entity (e.g., company, institution, organization)?
The dataset has been created by the Language Technologies unit (LangTech) of the Barcelona Supercomputing Center - Centro Nacional de
Supercomputación (BSC-CNS), which aims to advance the field of natural language processing through cutting-edge research and development
and the use of HPC. In particular, it was created by the unit's data team, the main contributors being Javier Saiz, Ferran Espuña, and
Jorge Palomar.
However, the creation of the dataset would not have been possible without the collaboration of a large number of collaborators, partners,
and public institutions, which can be found in detail in the acknowledgements.
Who funded the creation of the dataset? If there is an associated grant, please provide the name of the grantor and the grant name and number.
This work/research has been promoted and financed by the Government of Catalonia through the Aina project.
Composition
What do the instances that comprise the dataset represent (e.g., documents, photos, people, countries)? Are there multiple types of instances (e.g., movies, users, and ratings; people and interactions between them; nodes and edges)? Please provide a description.
The dataset consists entirely of text documents in various languages. Specifically, data was mainly sourced from the following databases and
repositories:
Common Crawl: Repository that holds website data and is run by the Common Crawl non-profit organization. It is updated monthly and is
distributed under the CC0 1.0 public domain license.
GitHub: Community platform that allows developers to create, store, manage, and share their code. Repositories are crawled and then
distributed with their original licenses, which may vary from permissive to non-commercial licenses.
Wikimedia: Database that holds the collection databases managed by the Wikimedia Foundation, including Wikipedia, Wikibooks, Wikinews,
Wikiquote, Wikisource, and Wikivoyage. It is updated monthly and is distributed under Creative Commons Attribution-ShareAlike License 4.0.
EurLex: Repository that holds the collection of legal documents from the European Union, available in all of the EU’s 24 official
languages and run by the Publications Office of the European Union. It is updated daily and is distributed under the Creative Commons
Attribution 4.0 International license.
Other repositories: Specific repositories were crawled under permission for domain-specific corpora, which include academic, legal,
and newspaper repositories.
We provide a complete list of dataset sources at the end of this section.
How many instances are there in total (of each type, if appropriate)?
The dataset contains a diverse range of instances across multiple languages, with notable adjustments for certain languages. English
represents the largest portion, accounting for 39.08% of the total data. Spanish was upsampled by a factor of 2, bringing its share to 16.59%,
while Catalan (1.84%), Basque (0.26%), and Galician (0.36%) were also upsampled by 2. On the other hand, code-related data was downsampled
by half, making up 6.42% of the total. Other prominent languages include French (6.59%), Russian (5.39%), German (4.25%), and Hungarian
(3.93%), with several additional languages contributing between 1% and 2%, and smaller portions represented by a variety of others.
Does the dataset contain all possible instances or is it a sample (not necessarily random) of instances from a larger set? If the dataset is a sample, then what is the larger set? Is the sample representative of the larger set (e.g., geographic coverage)? If so, please describe how this representativeness was validated/verified. If it is not representative of the larger set, please describe why not (e.g., to cover a more diverse range of instances, because instances were withheld or unavailable).
The dataset is a sample from multiple sources, with different weights based on the primary language of the content: Spanish, Catalan,
Basque, and Galician content was upsampled by a factor of two, while programming languages were downsampled by a factor of half. Other
sources were sampled in proportion to their occurrence.
What data does each instance consist of? “Raw” data (e.g., unprocessed text or images) or features? In either case, please provide a description.
Each instance consists of a text document processed for deduplication, language identification, and source-specific filtering. Some
documents required optical character recognition (OCR) to extract text from non-text formats such as PDFs.
Is there a label or target associated with each instance? If so, please provide a description.
Each instance is labeled with a unique identifier, the primary language of the content, and the URL for web-sourced instances. Additional
labels were automatically assigned to detect specific types of content —harmful or toxic content— and to assign preliminary indicators of
undesired qualities —very short documents, high density of symbols, etc.— which were used for filtering instances.
Is any information missing from individual instances? If so, please provide a description, explaining why this information is missing (e.g., because it was unavailable). This does not include intentionally removed information, but might include, e.g., redacted text.
No significant information is missing from the instances.
Are relationships between individual instances made explicit (e.g., users’ movie ratings, social network links)? If so, please describe how these relationships are made explicit.
Instances are related through shared metadata, such as source and language identifiers.
Are there recommended data splits (e.g., training, development/validation, testing)? If so, please provide a description of these splits, explaining the rationale behind them.
The dataset is split randomly into training, validation, and test sets.
Are there any errors, sources of noise, or redundancies in the dataset? If so, please provide a description.
Despite removing duplicated instances within each source, redundancy remains at the paragraph and sentence levels, particularly in
web-sourced instances where SEO techniques and templates contribute to repeated textual patterns. Some instances may also be duplicated
across sources due to format variations.
Is the dataset self-contained, or does it link to or otherwise rely on external resources (e.g., websites, tweets, other datasets)? If it links to or relies on external resources, a) are there guarantees that they will exist, and remain constant, over time; b) are there official archival versions of the complete dataset (i.e., including the external resources as they existed at the time the dataset was created); c) are there any restrictions (e.g., licenses, fees) associated with any of the external resources that might apply to a dataset consumer? Please provide descriptions of all external resources and any restrictions associated with them, as well as links or other access points, as appropriate.
The dataset is self-contained and does not rely on external resources.
Does the dataset contain data that might be considered confidential (e.g., data that is protected by legal privilege or by doctor–patient confidentiality, data that includes the content of individuals’ non-public communications)? If so, please provide a description.
The dataset does not contain confidential data.
Does the dataset contain data that, if viewed directly, might be offensive, insulting, threatening, or might otherwise cause anxiety? If so, please describe why. If the dataset does not relate to people, you may skip the remaining questions in this section.
The dataset includes web-crawled content, which may overrepresent pornographic material across languages (Kreutzer et al., 2022). Although
pre-processing techniques were applied to mitigate offensive content, the heterogeneity and scale of web-sourced data make exhaustive
filtering challenging, which makes it next to impossible to identify all adult content without falling into excessive filtering, which may
negatively influence certain demographic groups (Dodge et al., 2021).
Does the dataset identify any subpopulations (e.g., by age, gender)? If so, please describe how these subpopulations are identified and provide a description of their respective distributions within the dataset.
The dataset does not explicitly identify any subpopulations.
Is it possible to identify individuals (i.e., one or more natural persons), either directly or indirectly (i.e., in combination with other data) from the dataset? If so, please describe how.
Web-sourced instances in the dataset may contain personally identifiable information (PII) that is publicly available on the Web, such as
names, IP addresses, email addresses, and phone numbers. While it would be possible to indirectly identify individuals through the
combination of multiple data points, the nature and scale of web data makes it difficult to parse such information. In any case, efforts are
made to filter or anonymize sensitive data during pre-processing, but some identifiable information may remain in the dataset.
Does the dataset contain data that might be considered sensitive in any way? If so, please provide a description.
Given that the dataset includes web-sourced content and other publicly available documents, instances may inadvertently reveal financial
information, health-related details, or forms of government identification, such as social security numbers (Subramani et al., 2023),
especially if the content originates from less-regulated sources or user-generated platforms.
Collection Process
How was the data collected?
This dataset is constituted by combining several sources, whose acquisition methods can be classified into three groups:
Web-sourced datasets with some preprocessing available under permissive license (p.e. Common Crawl).
Domain-specific or language-specific raw crawls (p.e. Spanish Crawling).
Manually curated data obtained through collaborators, data providers (by means of legal assignment agreements) or open source projects
(p.e. CATalog).
What mechanisms or procedures were used to collect the data? How were these mechanisms or procedures validated?
According to the three groups previously defined, these are the mechanisms used in each of them:
Open direct download. Validation: data integrity tests.
Ad-hoc scrapers or crawlers. Validation: software unit and data integrity tests.
Direct download via FTP, SFTP, API or S3. Validation: data integrity tests.
If the dataset is a sample from a larger set, what was the sampling strategy?
The sampling strategy was to use the whole dataset resulting from the filtering explained in the ‘preprocessing/cleaning/labelling’ section,
with the particularity that an upsampling of 2 (i.e. twice the probability of sampling a document) was performed for the co-official
languages of Spain (Spanish, Catalan, Galician, Basque), and a downsampling of 1/2 was applied for code (half the probability of sampling a
code document, evenly distributed among all programming languages).
Who was involved in the data collection process and how were they compensated?
This data is generally extracted, filtered and sampled by automated processes. The code required to run these processes has been developed
entirely by members of the LangTech data team, or otherwise obtained from open-source software. Furthermore, there has been no monetary
consideration for acquiring data from suppliers.
Over what timeframe was the data collected? Does this timeframe match the creation timeframe of the data associated with the instances? If not, please describe the timeframe in which the data associated with the instances was created.
Data were acquired and processed from April 2023 to April 2024. However, as mentioned, much data has been obtained from open projects such
as Common Crawl, which contains data from 2014, so it is the end date (04/2024) rather than the start date that is important.
Were any ethical review processes conducted? If so, please provide a description of these review processes, including the outcomes, as well as a link or other access point to any supporting documentation.
No particular ethical review process has been carried out as the data is mostly open and not particularly sensitive. However, we have an
internal evaluation team and a bias team to monitor ethical issues. In addition, we work closely with ‘Observatori d'Ètica en Intel·ligència
Artificial’ (OEIAC) and ‘Agencia Española de Supervisión de la Inteligencia Artificial’ (AESIA) to audit the processes we carry out from an
ethical and legal point of view, respectively.
Preprocessing
Was any preprocessing/cleaning/labeling of the data done? If so, please provide a description. If not, you may skip the remaining questions in this section.
Instances of text documents were not altered, but web-sourced documents were filtered based on specific criteria along two dimensions:
Quality: documents with a score lower than 0.8, based on undesired qualities, such as documents with low number of lines, very short
sentences, presence of long footers and headers, and high percentage of punctuation, obtained through CURATE (Palomar-Giner et al., 2024)
were filtered out.
Harmful or adult content: documents originating from Colossal OSCAR were filtered using LLM-Datasets (Ostendorff et al., 2024) based on
the perplexity from a language model (‘harmful_pp’ field) provided by the Ungoliant pipeline (Abadji et al., 2021).
Was the “raw” data saved in addition to the preprocessed/cleaned/labeled data? If so, please provide a link or other access point to the “raw” data.
The original raw data was not kept.
Is the software that was used to preprocess/clean/label the data available? If so, please provide a link or other access point.
Yes, the preprocessing and filtering software is open-sourced. The CURATE pipeline was used for Spanish Crawling and CATalog,
and the Ungoliant pipeline was used for the OSCAR project.
Uses
Has the dataset been used for any tasks already? If so, please provide a description.
Pre-train the Salamandra model family.
What (other) tasks could the dataset be used for?
The data can be used primarily to pre-train other language models, which can then be used for a wide range of use cases. The dataset could
also be used for other tasks such as fine-tuning language models, cross-lingual NLP tasks, machine translation, domain-specific text
generation, and language-specific data analysis.
Is there anything about the composition of the dataset or the way it was collected and preprocessed/cleaned/labeled that might impact future uses? Is there anything a dataset consumer could do to mitigate these risks or harms?
Web-crawled content is over-represented with standard language varieties, impacting language model performance for minority languages.
Language diversity in data is crucial to avoid bias, especially in encoding non-standard dialects, preventing the exclusion of demographic
groups. Moreover, despite legal uncertainties in web-scraped data, we prioritize permissive licenses and privacy protection measures,
acknowledging the challenges posed by personally identifiable information (PII) within large-scale datasets. Our ongoing efforts aim to
address privacy concerns and contribute to a more inclusive linguistic dataset.
Are there tasks for which the dataset should not be used?
Distribution
Will the dataset be distributed to third parties outside of the entity on behalf of which the dataset was created? If so, please provide a description.
The dataset will not be released or distributed to third parties. Any related question to distribution is omitted in this section.
Maintenance
Who will be supporting/hosting/maintaining the dataset?
The dataset will be hosted by the Language Technologies unit (LangTech) of the Barcelona Supercomputing Center (BSC). The team will ensure
regular updates and monitor the dataset for any issues related to content integrity, legal compliance, and bias for the sources they are
responsible for.
How can the owner/curator/manager of the dataset be contacted?
The data owner may be contacted with the email address langtech@bsc.es.
Will the dataset be updated?
The dataset will not be updated.
If the dataset relates to people, are there applicable limits on the retention of the data associated with the instances? If so, please describe these limits and explain how they will be enforced.
The dataset does not keep sensitive data that could allow direct identification of individuals, apart from the data that is publicly
available in web-sourced content. Due to the sheer volume and diversity of web data, it is not feasible to notify individuals or manage data
retention on an individual basis. However, efforts are made to mitigate the risks associated with sensitive information through
pre-processing and filtering to remove identifiable or harmful content. Despite these measures, vigilance is maintained to address potential
privacy and ethical issues.
Will older versions of the dataset continue to be supported/hosted/maintained? If so, please describe how. If not, please describe how its obsolescence will be communicated to dataset consumers.
Since the dataset will not be updated, only the final version will be kept.
If others want to extend/augment/build on/contribute to the dataset, is there a mechanism for them to do so?
The dataset does not allow for external contributions.
Finetuning Data
This instruction-tuned variant has been trained with a mixture of 276k English, Spanish, and Catalan multi-turn instructions gathered from open datasets:
Dataset
ca
en
es
alpaca-cleaned
-
50,000
-
aya-dataset
-
3,944
3,854
CoQCat
4,797
-
-
databricks-dolly-15k
-
15,011
-
dolly-3k-ca
3,232
-
-
flores-instr
1,994
1,994
3,988
MentorCA
7,122
-
-
MentorES
-
-
7,122
no-robots
-
9,499
-
oasst-ca
2,518
-
-
oasst2
750
31,086
15,438
open-orca
-
50,000
-
RagMultilingual
16,043
14,997
11,263
tower-blocks
-
19,895
2,000
Total
36,456
196,426
43,665
Evaluation
Gold-standard benchmarks
Evaluation is done using the Language Model Evaluation Harness (Gao et al., 2024). We evaluate on a set of tasks taken from SpanishBench, CatalanBench, BasqueBench and GalicianBench. These benchmarks include both new and existing tasks and datasets. Given that this is an instructed model, we add LM Evaluation Harness's native feature of chat-template to the setup. In the tables below, we include the results in a selection of evaluation datasets that represent model's performance across a variety of tasks within these benchmarks.
We only use tasks that are either human generated, human translated, or with a strong human-in-the-loop (i.e., machine translation followed by professional revision or machine generation followed by human revision and annotation). This is the reason behind the variety in number of tasks reported across languages. As more tasks that fulfill these requirements are published, we will update the presented results. We also intend to expand the evaluation to other languages, as long as the datasets meet our quality standards.
During the implementation of the evaluation we observed a series of issues worth considering when replicating and interpreting the results presented. These issues include ≈1.5% variances in performance in some tasks depending on the version of the transformers library used, and depending on the use (or lack of use) of tensor parallelism when loading a model. When implementing existing tasks, we carry out a comprehensive quality evaluation of the dataset, the Harness task itself, and what kind of input models see during evaluation. Our implementation (see links above) addresses multiple existing problems such as errors in datasets and prompts, and lack of pre-processing. All this means that results will vary if using other Harness implementations, and may slightly vary depending on the replication setup.
It should be noted that these results are subject to all the drawbacks of every current gold-standard evaluation, and that the figures do not fully represent the models capabilities and potential. We thus advise caution when reading and interpreting the results.
A full list of results compared to other baselines, a discussion of the model's performance across tasks and its implications, and details regarding problem-solving with task implementation will soon be available in the technical report.
All results reported below are on a 0-shot setting.
Spanish
Category
Task
Metric
Result
Commonsense Reasoning
xstorycloze_es
acc
69.29
NLI
wnli_es
acc
45.07
xnli_es
acc
51.49
Paraphrasing
paws_es
acc
59.4
QA
xquad_es
acc
43.82
Translation
flores_es
bleu
22.98
Catalan
Category
Task
Metric
Result
Commonsense Reasoning
copa_ca
acc
81.2
xstorycloze_ca
acc
70.68
NLI
wnli_ca
acc
50.7
xnli_ca
acc
55.14
Paraphrasing
parafraseja
acc
65.18
paws_ca
acc
62.95
QA
arc_ca_easy
acc
64.98
arc_ca_challenge
acc
41.89
openbookqa_ca
acc
35.2
piqa_ca
acc
69.53
siqa_ca
acc
48.62
Translation
flores_ca
bleu
28.65
Basque
Category
Task
Metric
Result
Commonsense Reasoning
xcopa_eu
acc
61.6
xstorycloze_eu
acc
61.15
NLI
wnli_eu
acc
45.07
xnli_eu
acc
46.81
QA
eus_exams
acc
39.09
eus_proficiency
acc
36.93
eus_trivia
acc
46.94
Reading Comprehension
eus_reading
acc
45.45
Translation
flores_eu
bleu
14.89
Galician
Category
Task
Metric
Result
Paraphrasing
parafrases_gl
acc
55.44
paws_gl
acc
56.55
QA
openbookqa_gl
acc
38.4
Translation
flores_gl
bleu
27.03
LLM-as-a-judge
We use Prometheus-2 8x7B as a judge to evaluate the responses of the model. Tasks are created from existing multilingual evaluation datasets covering the same categories as the ones measured in our gold-standard benchmarks. We randomly select a subset of 250 instances per language from the test set of each source dataset. To evaluate the responses of our model, we use task-specific criteria developed in-house for the LLM-judge to use. Each criterion is measured either as a 5-point Likert scale or as a binary task depending on the idiosyncrasy of the task and criterion.
Prompts for each task are created in various ways to score the model's robustness in addition to these criteria. This is done by presenting the same source instance within three different prompts. We then calculate the variance between the scores assigned by the LLM-judge to our model's responses to the three prompt styles and average it across all instances. Prompts are human translated to all languages measured. We do not provide the LLM-judge with a reference answer.
The judge prompt we use during evaluation is the same used to fine tune the Prometheus-2 family. We keep the judge prompt and criteria used to present the LLM-judge with the task prompts and model responses in English for evaluation across languages. The judge prompt used is:
python
1"You are a fair judge assistant tasked with providing clear, objective feedback based on specific criteria, ensuring each assessment reflects the absolute standards setfor performance.23###Task Description:4An instruction (might include an Input inside it), a response to evaluate,and a score rubric representing a evaluation criteria are given.51. Write a detailed feedback that assess the quality of the response strictly based on the given score rubric,not evaluating in general.62. After writing a feedback, write a score that is an integer between {a}and{b}. You should refer to the score rubric.73. The output format should look as follows: \"Feedback:(write a feedback for criteria)[RESULT](an integer number between {a}and{b})\"
84. Please do not generate any other opening, closing,and explanations.910###The instruction to evaluate:11{input}1213###Response to evaluate:14{prediction}1516###Score Rubrics:17{criteria}1819###Feedback:"
As an example, prompts for the Math task in English are based on instances from MGSM, and each instance is presented within these prompts:
python
1"en":[2("I need help with this math problem: \"","\" Give me the answer step by step and also the final result separately."),3("Can you please help me answer this? \"","\" Explain the answer and give me the final result as well. Thanks."),4("Help me with this problem: \"","\" I need the answer explained and the final result separately.")5]
This task is then evaluated by the LLM-judge using two criteria, reasoning capability (5-point Likert) and mathematical correctness (binary):
python
1reasoning_capability_criteria ={2"reasoning_capability":"""
3[Does the model's answer demonstrate reasoning capability?]
4Score 1: The answer demonstrates poor reasoning, with illogical arguments or conclusions that do not follow from the provided information.
5Score 2: The answer shows weak reasoning, with some logical connections but also contains significant flaws or gaps in the argumentation.
6Score 3: The answer demonstrates adequate reasoning, with generally logical arguments, but may have minor flaws or a lack of depth in the reasoning process.
7Score 4: The answer shows strong reasoning, with well-structured arguments and conclusions that logically follow from the information provided.
8Score 5: The answer demonstrates exceptional reasoning, with clear, coherent, and insightful arguments that are logically sound and well-supported by the information provided."""9}1011mathematical_correctness_binary_criteria ={12"mathematical_correctness_binary":"""
13[Is the model's answer mathematically correct?]
14Score 0: The answer contains mathematical errors that render the solution incorrect or unreliable.
15Score 1: The answer is mathematically correct, with accurate calculations and appropriate use of mathematical concepts."""16}
Multilingual results
Here, we present results for seven categories of tasks in Spanish, Catalan, Basque, Galician, and English. Results are presented for each task, criterion and language. Criteria with a (B) after their name are binary criteria (i.e., numbers go from 0 to 1, where 1 is best). The rest of the criteria are measured using a 5-point Likert scale, where 5 is best. The first number of the pair of numbers separated by / shows the average score for the criterion (and language). The second number of each pair is the robustness score, where numbers closer to 0 mean that the model generates similar responses when comparing the three prompt varieties for a single instance.
Further details on all tasks and criteria, a full list of results compared to other baselines, a discussion of the model's performance across tasks and its implications, and details regarding problem-solving with task implementation will soon be available in the technical report.
Ethical Considerations and Limitations
We examine the presence of undesired societal and cognitive biases present in this model using different benchmarks. For societal biases,
we test performance using the BBQ dataset (Parrish et al., 2022) in the original English and the Regard dataset (Sheng et al., 2019).
We report that while performance is high (accuracies around 0.8 depending on the social category) in disambiguated settings,
the model performs very poorly in ambiguous settings, which indicates the presence of societal biases that need to be further addressed in post-training phases.
Our cognitive bias analysis focuses on positional effects in 0-shot settings, and majority class bias in few-shot settings.
For positional effects, we leverage the ARC Multiple Choice Question dataset (Clark et al., 2018). We observe significant,
but relatively weak primacy effects, whereby the model shows a preference for answers towards the beginning of the list of provided answers.
We measure effects of majority class effects in few-shot settings using SST-2 (Socher et al., 2013). We again detect significant effects,
with a small effect size. This suggests that the model is relatively robust against the examined cognitive biases.
We highlight that our analyses of these biases are by no means exhaustive and are limited by the relative scarcity of adequate resources
in all languages present in the training data. We aim to gradually extend and expand our analyses in future work.
These results can be expected from a model that has undergone only a preliminary instruction tuning.
These tests are performed in order to show the biases the model may contain. We urge developers to take
them into account and perform safety testing and tuning tailored to their specific applications of the model.
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) 2024 by Language Technologies Unit, Barcelona Supercomputing Center.
Funding
This work has been promoted and financed by the Government of Catalonia through the Aina Project.
This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU
within the framework of ILENIA Project with reference 2022/TL22/00215337.
Acknowledgements
This project has benefited from the contributions of numerous teams and institutions, mainly through data contributions, knowledge transfer or technical support.
In Catalonia, many institutions have been involved in the project. Our thanks to Òmnium Cultural, Parlament de Catalunya, Institut d'Estudis Aranesos, Racó Català, Vilaweb, ACN, Nació Digital, El món and Aquí Berguedà.
At national level, we are especially grateful to our ILENIA project partners: CENID, HiTZ and CiTIUS for their participation. We also extend our genuine gratitude to the Spanish Senate and Congress, Fundación Dialnet, Fundación Elcano and the ‘Instituto Universitario de Sistemas Inteligentes y Aplicaciones Numéricas en Ingeniería (SIANI)’ of the University of Las Palmas de Gran Canaria.
At the international level, we thank the Welsh government, DFKI, Occiglot project, especially Malte Ostendorff, and The Common Crawl Foundation, especially Pedro Ortiz, for their collaboration. We would also like to give special thanks to the NVIDIA team, with whom we have met regularly, specially to: Ignacio Sarasua, Adam Henryk Grzywaczewski, Oleg Sudakov, Sergio Perez, Miguel Martinez, Felipes Soares and Meriem Bendris. Their constant support has been especially appreciated throughout the entire process.
Their valuable efforts have been instrumental in the development of this work.
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.
