In recent years Large Language Models have improved the state of the art on several natural language processing tasks. However, their availability is frequently restricted to paid API services, posing challenges for researchers in conducting extensive investigations. On the other hand, while some open-source models have been proposed by the community, they are typically English-centric or multilingual without a specific adaptation for the Italian language. In an efort to democratize the available and open resources for the Italian language, in this paper we introduce Camoscio: a language model specifically tuned to follow users' prompts in Italian. Specifically, we finetuned the smallest variant of LLaMA (7b) with LoRA on a corpus of instruction prompts translated to Italian via ChatGPT. Results indicate that the model's zero-shot performance on various downstream tasks in Italian competes favorably with existing models specifically finetuned for those tasks. All the artifacts (code, dataset, model) are released to the community at the following url: https://github.com/teelinsan/camoscio
CEUR ceur-ws.org
In recent years, Large Language Models (LLMs) have
made remarkable advancements in the field of natural
language processing, demonstrating state-of-the-art
performance on various tasks [
On the other hand, several open-source models1 have
been proposed as an alternative to closed models [
CEUR
Workshop Proceedings (CEUR-WS.org) 1Actual openness depends on the model license.
2Less than 4.5% of training data comes from Wikipedia in 20 nEvelop-O (E. Rodolà) CEUR htp:/ceur-ws.org
ISN1613-073
diferent languages, including Italian.
of raw textual data, while it has been shown that a
second training step of instruction-tuning is crucial to
increase downstream performance [
• We introduce an instruction-tuning dataset for the Italian language, stemming from the Stanford
• We train Camoscio on this dataset and evaluate its zero-shot performance on several downstream tasks for the Italian language (NewsSum
Large language models have emerged as a general class
of models capable of performing a wide range of tasks
without explicit finetuning by just leveraging in-context
examples [
In the context of the Italian language, the availability of pre-trained language models is currently limited; generic multipurpose LMs are almost nonexistent. Notable mentions include: AlBERTo [17], an Italian version of BERT [18] trained on Italian tweets from TWITA [19]; GePpeTto [20], a version of GPT-2 base (117 million parameters) finetuned using Italian Wikipedia and the ItWac corpus [21]; IT5 [22] a T5 model tailored for Italian using a refined version of the mC4 corpus [ 23]; and BART-IT [24], an Italian variant of BART [25] trained on the same mixture of data as IT5. Concurrently to our work, Bacciu et al. [26] proposed Fauno, an Italian version of Baize [27] that is a LM trained on a corpus of self-chat performed by ChatGPT. Compared to our work, their approach is tailored to develop a conversational agent for the Italian language. After our work, Michael [28] released on their GitHub repository an instructiontuned version of LLaMA on a translation to Italian of the GPT-4-LLM dataset [29].
the Italian language, we stem from the Stanford Alpaca
dataset [
imi.”, Input: “Luce”, Output: “Scuro, pesante, denso”. 3.1. Dataset Clearly the translation is not always perfect, but it is a fast-and-cheap method to bootstrap a noisy instructionStanford Alpaca is an instruction-tuning dataset con- tuning dataset for the Italian language. structed using the self-instruct method [31]. Specifically, the authors started with a set of 175 human-written instruction-output pairs from the original self-instruct 3.2. Training & Prompting paper3 and used them as in-context examples to prompt OpenAI text-davinci-003. A total of 52.000 novel examples are generated with this technique. Each example includes an instruction, in natural English language, the answer (output), and optionally an additional context (input) for some datapoints (e.g., a short paragraph for question answering). Figure 1 shows diferent types of instructions in the dataset.
At inference time, the same prompt is used to gener- or finetuning to this kind of articles, although we recogate the answer. Only the text generated after “[...] ### nize it might be interesting to analyze this in a few-shot Risposta:” is used as final output. We sample from the setting). We describe in the next paragraphs the three model using top-p sampling [37] with a temperature of datasets used for the evaluation. 0.2, = 0.75 , = 40 , and beam search with 4 beams.
We refer to Appendix A for the additional implementation details.
Currently, there is a very limited availability of datasets for a solid evaluation of the broad capabilities these general-purpose models possess. This is true for English but especially for the Italian language, although the community is moving towards this direction [38]. To evaluate our model we decided to follow the same evaluation protocol proposed in Sarti and Nissim [22]. Compared to their approach, we do not perform any training on the downstream tasks, i.e., we perform just the evaluation on the test set in a zero-shot fashion by providing to the model a textual description of the task (e.g., “Riassumi Question Answering. To assess the model perforil seguente articolo”). We compared the performance of mance on extractive question answering, we used the our model on standard Italian benchmarks for summa- SQuAD-IT dataset [32]. This dataset is composed of sets rization (NewsSum-IT), question answering (SQuAD-IT), of paragraphs, questions, and answers derived from the and style transfer (XFORMAL IT). original SQuAD dataset [42] via machine translation and
Compared to Sarti and Nissim [22], we do not include subsequent filtering of problematic instances. As for the
the Wikipedia for Italian Text Summarization (WITS) previous datasets, we used just the test split for zero-shot
corpus [39] since Wikipedia is included in the original evaluation. The model is asked to generate an answer
training corpus of LLaMA [
summarization capabilities of Camoscio using the dataset NewSum-IT proposed by Sarti and Nissim [22]. This dataset is obtained by merging two newspaper sources (“Fanpage.it” and “Il Post”) scraped by the Applied Recognition Technology Laboratory4 and available on the Hugging Face Hub [41]. We used only the test split for the zero-shot evaluation and asked the model to generate an answer given the instruction “Dopo aver letto il testo qui sotto, riassumilo adeguatamente.” provided in the textual prompt and the news text provided as input (complete prompt as explained in §3.2). We use the same evaluation metrics of Sarti and Nissim [22] and report the average across the two newspapers as in their work. 4.2. Results and Discussion evaluated the generated answers using the script from even with a correct output but diferent wording. To acSarti and Nissim [22]. Furthermore, we also used an ad- count for these variations, we used an approach similar ditional metric “ChatGPT Exact Match” to better assess to Zheng et al. [47] that leverages an external LM (in our the performance. We explain this metric in the following case gpt-3.5-turbo) to judge whether the answer provided subsection “Evaluation Metrics”. by a model is correct (1) or not (0) given the question and the ground-truth answer. We refer to this metric as Exact Match via ChatGPT (EM-GPT) and explain it with additional details in Appendix B.
Formality Style Transfer. We assess the style transfer capabilities of Camoscio using the Italian subset of the XFORMAL dataset [43], hereafter referred to as XFORMAL-IT. The dataset consists of forum messages from the GYAFC corpus [44] automatically translated covering several topics (entertainment, music, family, and relationships). The test set is constructed by using crowdworkers via Amazon Mechanical Turk to collect formal-informal pairs directly in Italian. The model is evaluated in both style transfer directions (Formal to Informal and Informal to Formal). We use only the test split for the zero-shot evaluation and ask the model to generate an answer given the instruction “Dato il seguente testo scritto in modo formale, riscrivilo in modo informale.” and vice versa according to the style transfer direction.
Camoscio compared to other methods used in the literature. We observe that the metrics commonly used for the task (Exact Match and F1) are very low compared to all the other models. Although this is generally expected since we are comparing trained models with an untrained one, the exact match score is suspiciously low.
Looking at the output responses, we noted that Camoscio produces correct but wordy answers (e.g., “La crisi petrolifera del 1973 è iniziata nell’ottobre 1973.” instead of “ottobre 1973”) making the system to perform bad on this score despite the fact that it produces correct answers. 4.1. Evaluation Metrics Since all the other systems are trained on the datasets, We use the same evaluation protocol and scripts of they are aligned with the expected target distribution Sarti and Nissim [22]. Specifically, for evaluating lexical and the exact match metric is an efective choice. Nevmatches, we rely on the language-independent ROUGE ertheless, when it comes to the zero-shot configuration metric proposed by Lin [45] in the variants unigram (R1), in Camoscio, this conventional metric fails to accurately bigram (R2), and Longest Common Subsequence (RL). To capture the true performance of the task. gauge semantic correspondence, we employ the trained To this end, we evaluated the model also with standard BERTScore metric [46] with a widely used BERT model evaluation metrics for generative models (R1, R2, RL, BS). pre-trained on Italian5 and the same baseline scores as However, we also observe in this case low scores despite Sarti and Nissim [22]. Following previous works, for eval- the fact that a qualitative examination of the provided uating the Question-Answering task we employ exact- answers suggests an overall higher quality. This is posmatch (EM) and F1-score (F1). However, since Camoscio sibly due to the diferent lengths between the produced is not trained on the output distribution of the question- answers (long) and the ground truth (short) and reinanswering dataset, these metrics will fail to assess the forces the necessity of developing a more precise metric correctness of the output since the EM will count as zero to accurately gauge task performance.
For this purpose, we used instead the metric Exact
R1 .377 “Questo articolo non è più commentabile. Abbonati al Post per commentare le altre notizie.”. We found this behavior peculiar, considering that the input prompt never mentions that articles were collected from “Il Post”. Looking carefully at the dataset, we found that these documents were not correctly scraped from the web and the model was simply verbatim copying the input error message above (i.e., the message is already summarized and the model is proposing it as-is).
Open-ended generation. We show here some qualitative results on the task of open-ended generation. Compared to the other tasks above, in this task we provide the model with just the instruction without any specific Match via ChatGPT explained in §4.1. This metric shows input. As a tribute to GePpeTto [20], the first generative that the actual zero-shot performance of Camoscio is in language model for Italian, we pose to the model the line with the other trained models (.576) and it is also following instruction: “Riassumi la storia di Pinocchio”. way higher compared to the original EM metric (.077), The model answered: “La storia di Pinocchio racconta di confirming the need for another type of metric to evalu- un bambino di nome Pinocchio che vive con i suoi genitori ate the task in the zero-shot setting. Results also show in un piccolo villaggio. Un giorno, Pinocchio incontra un that the EM-GPT metric of trained models correlates well burattino magico che gli dona la possibilità di diventare with the existing EM metric, even though with a little un vero bambino. Pinocchio inizia una serie di avventure marginal diference. This suggests that this metric could mentre cerca di diventare un vero bambino. Durante il serve as an approximate estimation of the model’s ac- suo viaggio, Pinocchio fa amicizia con molti personaggi e tual performance, although it might be subject to bias impara molte lezioni di vita. Alla fine, Pinocchio raggiunge according to the model used for estimation. il suo obiettivo e diventa un vero bambino”. At first glance the answer might seem correct, it is grammatically accuStyle Transfer & Summarization. Tables 2 and 3 rate and fluent in Italian. However, knowing the story, show results respectively for the formality style transfer the model makes several factual errors like characterand news summarization task. We can observe that the izing Pinocchio as a kid (instead of a wooden puppet) zero-shot performance of Camoscio in both tasks is com- living with parents (instead of Geppetto) and introducpetitive with trained models. According to the model and ing a nonexistent magic puppet. Overall this example training dataset, these latter might achieve slightly better highlights that, although promising, these models have scores at the expense of a less generalist model. Looking well-known limitations like hallucinations, factual errors, at the qualitative results, we note however that the sum- and several kinds of biases [48, 49, 50]. Consequently, marization task on “Il Post” and “Fanpage” is afected by it is essential to exercise caution when utilizing them, some common failure cases. keeping these limitations in mind. Failure Cases. The most common failure case consists of the model not producing an answer at all after the input prompt (4.93% of cases on “Il Post” and 21.16% cases on “Fanpage”). We think that it might be due to the input document of these examples being too long and out of distribution compared to the training documents seen in the instruction-tuning dataset (max length 256 tokens). This might confuse the model and trigger the generation of the end-of-sentence token. Other failure cases include the model responding with a copy of the instruction prompt. For example, we found that in 0.75% of the cases in the “Il Post” split of NewSum-IT, when asked to summarize the article the model responded with
7https://huggingface.co/ARTeLab/ mbart-summarization-fanpage
In this paper, we introduced Camoscio, a 7 billion instruction-tuned model for the Italian language, together with its Italian instruction-tuning dataset. Results show that the zero-shot performance of Camoscio on several downstream tasks in Italian is competitive with existing models specifically finetuned for those tasks. Despite the known limitations of these kinds of models, this is a first step towards a generalist model capable of performing a wide range of tasks in Italian without explicit ifnetuning. This is particularly relevant especially in several domains where data is scarce or not available (e.g., medical domain). In an efort to democratize the available and open resources for the Italian language, we release all the artifacts (code, dataset, model) to the community.
formance competitive with existing finetuned models on three diferent tasks: summarization (NewsSumIT), question answering (SQuAD-IT), and style transfer (XFORMAL IT). However, it is unclear whether this is true also for other tasks, especially those out of training distribution of the instruction-tuning dataset (see Figure 1). Evaluating and thoroughly assessing the performance of these kinds of models is still an open research question. In addition to this, as already mentioned, the model suffers from common problems that afect language models such as hallucinations, factual errors, and several kinds of biases.
mentation issues with the tokenization and the training objective in the alpaca-lora repository and Gabriele Sarti for sharing datasets and evaluation protocols used in IT5.
We thank all the anonymous reviewers at CLiC-it 2023 for the feedback provided. mation Retrieval, SIGIR ’23, Association for Com- putational Linguistics, Online, 2021, pp. 483–498. puting Machinery, New York, NY, USA, 2023, p. URL: https://aclanthology.org/2021.naacl-main.41. 2619–2628. URL: https://doi.org/10.1145/3539618. doi:10.18653/v1/2021.naacl- main.41. 3591930. doi:10.1145/3539618.3591930. [24] M. La Quatra, L. Cagliero, Bart-it: An eficient [16] A. Santilli, S. Severino, E. Postolache, V. Maiorca, sequence-to-sequence model for italian text sumM. Mancusi, R. Marin, E. Rodola, Accelerating marization, Future Internet 15 (2023). URL: https: transformer inference for translation via parallel //www.mdpi.com/1999-5903/15/1/15. doi:10.3390/ decoding, in: Proceedings of the 61st Annual Meet- fi15010015. ing of the Association for Computational Linguis- [25] M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, tics (Volume 1: Long Papers), Association for Com- A. Mohamed, O. Levy, V. Stoyanov, L. Zettleputational Linguistics, Toronto, Canada, 2023, pp. moyer, BART: Denoising sequence-to-sequence 12336–12355. URL: https://aclanthology.org/2023. pre-training for natural language generation, transacl-long.689. lation, and comprehension, in: Proceedings of the [17] M. Polignano, P. Basile, M. De Gemmis, G. Semeraro, 58th Annual Meeting of the Association for ComV. Basile, et al., Alberto: Italian bert language under- putational Linguistics, Association for Computastanding model for nlp challenging tasks based on tional Linguistics, Online, 2020, pp. 7871–7880. URL: tweets, in: CEUR Workshop Proceedings, volume https://aclanthology.org/2020.acl-main.703. doi:10. 2481, CEUR, 2019, pp. 1–6. 18653/v1/2020.acl- main.703. [18] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: [26] A. Bacciu, G. Trappolini, A. Santilli, E. Rodolà, F. SilPre-training of deep bidirectional transformers for vestri, Fauno: The italian large language model language understanding, in: Proceedings of the that will leave you senza parole!, in: F. M. Nar2019 Conference of the North American Chap- dini, N. Tonellotto, G. Faggioli, A. Ferrara (Eds.), ter of the Association for Computational Linguis- Proceedings of the 13th Italian Information Retics: Human Language Technologies, Volume 1 trieval Workshop (IIR 2023), Pisa, Italy, June 8(Long and Short Papers), Association for Com- 9, 2023, volume 3448 of CEUR Workshop Proceedputational Linguistics, Minneapolis, Minnesota, ings, CEUR-WS.org, 2023, pp. 9–17. URL: https: 2019, pp. 4171–4186. URL: https://aclanthology.org/ //ceur-ws.org/Vol-3448/paper-24.pdf.
N19-1423. doi:10.18653/v1/N19- 1423. [27] C. Xu, D. Guo, N. Duan, J. McAuley, Baize: An open[19] V. Basile, M. Lai, M. Sanguinetti, et al., Long-term source chat model with parameter-eficient tuning social media data collection at the university of on self-chat data, 2023. arXiv:2304.01196. turin, in: Proceedings of the Fifth Italian Confer- [28] Michael, Stambecco: Italian instruction-following ence on Computational Linguistics (CLiC-it 2018), llama model, https://github.com/mchl-labs/ CEUR-WS, 2018, pp. 1–6. stambecco, 2023. [20] L. D. Mattei, M. Cafagna, F. Dell’Orletta, M. Nissim, [29] B. Peng, C. Li, P. He, M. Galley, J. Gao, Instruction M. Guerini, Geppetto carves italian into a language tuning with gpt-4, arXiv preprint arXiv:2304.03277 model, in: J. Monti, F. Dell’Orletta, F. Tamburini (2023). (Eds.), Proceedings of the Seventh Italian Confer- [30] E. J. Wang, Alpaca-lora, https://github.com/tloen/ ence on Computational Linguistics, CLiC-it 2020, alpaca-lora, 2023.
Bologna, Italy, March 1-3, 2021, volume 2769 of [31] Y. Wang, Y. Kordi, S. Mishra, A. Liu, N. A. Smith, CEUR Workshop Proceedings, CEUR-WS.org, 2020. D. Khashabi, H. Hajishirzi, Self-instruct: Aligning URL: https://ceur-ws.org/Vol-2769/paper_46.pdf. language models with self-generated instructions, [21] M. Baroni, S. Bernardini, A. Ferraresi, E. Zanchetta, in: Proceedings of the 61st Annual Meeting of the The wacky wide web: a collection of very large Association for Computational Linguistics (Volume linguistically processed web-crawled corpora, Lan- 1: Long Papers), Association for Computational Linguage resources and evaluation 43 (2009) 209–226. guistics, Toronto, Canada, 2023, pp. 13484–13508. [22] G. Sarti, M. Nissim, It5: Large-scale text-to-text URL: https://aclanthology.org/2023.acl-long.754. pretraining for italian language understanding and [32] D. Croce, A. Zelenanska, R. Basili, Neural learngeneration, arXiv preprint arXiv:2203.03759 (2022). ing for question answering in italian, in: AI* IA [23] L. Xue, N. Constant, A. Roberts, M. Kale, R. Al- 2018–Advances in Artificial Intelligence: XVIIth Rfou, A. Siddhant, A. Barua, C. Rafel, mT5: International Conference of the Italian Association A massively multilingual pre-trained text-to-text for Artificial Intelligence, Trento, Italy, November transformer, in: Proceedings of the 2021 Con- 20–23, 2018, Proceedings 17, Springer, 2018, pp. ference of the North American Chapter of the 389–402.
Association for Computational Linguistics: Hu- [33] A. Scaiella, D. Croce, R. Basili, Large scale datasets man Language Technologies, Association for Com- for image and video captioning in italian, Italian Journal of Computational Linguistics 2 (2019) 49–60. A community library for natural language proURL: http://www.ai-lc.it/IJCoL/v5n2/IJCOL_5_2_3_ cessing, in: Proceedings of the 2021 Confer__scaiella_et_al.pdf. ence on Empirical Methods in Natural Language [34] C. Larcher, M. Piau, P. Finardi, P. Gengo, P. Espos- Processing: System Demonstrations, Association ito, V. Caridá, Cabrita: closing the gap for foreign for Computational Linguistics, Online and Punta languages, 2023. arXiv:2308.11878. Cana, Dominican Republic, 2021, pp. 175–184. [35] D. Croce, G. Brandi, R. Basili, Deep bidirectional URL: https://aclanthology.org/2021.emnlp-demo.21. transformers for italian question answering, in: arXiv:2109.02846.
R. Bernardi, R. Navigli, G. Semeraro (Eds.), Proceed- [42] P. Rajpurkar, J. Zhang, K. Lopyrev, P. Liang, Squad: ings of the Sixth Italian Conference on Computa- 100,000+ questions for machine comprehension of tional Linguistics, Bari, Italy, November 13-15, 2019, text, in: Proceedings of the 2016 Conference on volume 2481 of CEUR Workshop Proceedings, CEUR- Empirical Methods in Natural Language Processing, WS.org, 2019. URL: https://ceur-ws.org/Vol-2481/ 2016, pp. 2383–2392.
paper25.pdf. [43] E. Briakou, D. Lu, K. Zhang, J. Tetreault, Olá, bon[36] A. Riabi, T. Scialom, R. Keraron, B. Sagot, D. Sed- jour, salve! xformal: A benchmark for multilingual dah, J. Staiano, Synthetic data augmentation for formality style transfer, in: Proceedings of the 2021 zero-shot cross-lingual question answering, in: Conference of the North American Chapter of the Proceedings of the 2021 Conference on Empirical Association for Computational Linguistics: Human Methods in Natural Language Processing, 2021, pp. Language Technologies, 2021, pp. 3199–3216. 7016–7030. [44] S. Rao, J. Tetreault, Dear sir or madam, may i intro[37] A. Holtzman, J. Buys, L. Du, M. Forbes, Y. Choi, duce the gyafc dataset: Corpus, benchmarks and The curious case of neural text degeneration, in: metrics for formality style transfer, in: ProceedInternational Conference on Learning Representa- ings of the 2018 Conference of the North American tions, 2020. URL: https://openreview.net/forum?id= Chapter of the Association for Computational LinrygGQyrFvH. guistics: Human Language Technologies, Volume 1 [38] V. Basile, L. Bioglio, A. Bosca, C. Bosco, V. Patti, (Long Papers), 2018, pp. 129–140.
UINAUIL: A unified benchmark for Italian natural [45] C.-Y. Lin, ROUGE: A package for automatic evallanguage understanding, in: Proceedings of the uation of summaries, in: Text Summarization 61st Annual Meeting of the Association for Com- Branches Out, Association for Computational Linputational Linguistics (Volume 3: System Demon- guistics, Barcelona, Spain, 2004, pp. 74–81. URL: strations), Association for Computational Linguis- https://aclanthology.org/W04-1013. tics, Toronto, Canada, 2023, pp. 348–356. URL: [46] T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, https://aclanthology.org/2023.acl-demo.33. doi:10. Y. Artzi, Bertscore: Evaluating text generation with 18653/v1/2023.acl- demo.33. bert, in: International Conference on Learning [39] S. Casola, A. Lavelli, WITS: wikipedia for italian Representations, 2019.
text summarization, in: E. Fersini, M. Passarotti, [47] L. Zheng, W.-L. Chiang, Y. Sheng, S. Zhuang, V. Patti (Eds.), Proceedings of the Eighth Italian Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, Conference on Computational Linguistics, CLiC-it H. Zhang, J. E. Gonzalez, I. Stoica, Judging llm-as2021, Milan, Italy, January 26-28, 2022, volume 3033 a-judge with mt-bench and chatbot arena, 2023. of CEUR Workshop Proceedings, CEUR-WS.org, 2021. arXiv:2306.05685.
URL: https://ceur-ws.org/Vol-3033/paper65.pdf. [48] Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, [40] L. De Mattei, M. Cafagna, A. AI, F. Dell’Orletta, Y. J. Bang, A. Madotto, P. Fung, Survey of halM. Nissim, A. Gatt, Change-it@ evalita 2020: lucination in natural language generation, ACM Change headlines, adapt news, generate, Proceed- Comput. Surv. 55 (2023). URL: https://doi.org/10. ings of the Seventh Evaluation Campaign of Natural 1145/3571730. doi:10.1145/3571730. Language Processing and Speech Tools for Italian [49] E. M. Bender, T. Gebru, A. McMillan-Major, (EVALITA 2020) 2765 (2020). S. Shmitchell, On the dangers of stochastic par[41] Q. Lhoest, A. Villanova del Moral, Y. Jernite, rots: Can language models be too big?, in: ProA. Thakur, P. von Platen, S. Patil, J. Chaumond, ceedings of the 2021 ACM Conference on Fairness, M. Drame, J. Plu, L. Tunstall, J. Davison, M. Šaško, Accountability, and Transparency, FAccT ’21, AssoG. Chhablani, B. Malik, S. Brandeis, T. Le Scao, ciation for Computing Machinery, New York, NY, V. Sanh, C. Xu, N. Patry, A. McMillan-Major, USA, 2021, p. 610–623. URL: https://doi.org/10.1145/ P. Schmid, S. Gugger, C. Delangue, T. Matussière, 3442188.3445922. doi:10.1145/3442188.3445922. L. Debut, S. Bekman, P. Cistac, T. Goehringer, [50] E. Sheng, K.-W. Chang, P. Natarajan, N. Peng, The V. Mustar, F. Lagunas, A. Rush, T. Wolf, Datasets: woman worked as a babysitter: On biases in lan
Finetuning technique [
Il governo sta valutando l’estensione dell’obbligo di mascherina all’aperto a tutto il Paese.
Continuano ad aumentare i nuovi casi di coronavirus nel nostro Paese. Sono stati 2.800 i contagi registrati ieri: numeri che preoccupano il governo e che ricordano quelli delle fasi più critiche dell’emergenza. Domani l’esecutivo mT5-base generated text: si riunirà e valuterà se sia il caso di rendere più severe Il governo sta pensando di introdurre nuove regole le norme anti-contagio attualmente in vigore. Entro la anti-contagio per frenare la curva dei contagi. Il provvedprossima settimana si attende il nuovo Dpcm contenente imento è stato esteso a tutto il territorio nazionale, le misure di contrasto all’epidemia, mentre si valuta la mentre si valuta la proroga dello stato di emergenza fino proroga dello stato di emergenza fino al prossimo 31 al prossimo 31 gennaio 2021. Vediamo quindi quali sono gennaio 2021. Ma vediamo quindi quali sono queste le nuove regole che il governo sta pensando di introdurre nuove regole che il governo sta pensando di introdurre per frenare la curva dei contagi. per frenare la curva dei contagi. L’obbligo di portare la mascherina all’aperto, già introdotto nei giorni scorsi in Ground Truth: alcune zone, sarà esteso a tutto il territorio nazionale. Con i nuovi casi di coronavirus in continuo aumento, il Oltre quindi a confermare la necessità di indossare governo pensa a inasprire le regole anti-contagio per consempre il dispositivo di protezione nei luoghi chiusi, di trastare la difusione dell’infezione. Si va verso l’obbligo igienizzare frequentemente le mani e di rispettare le di indossare la mascherina anche all’aperto e la chiusura distanze di sicurezza e il divieto di assembramento, il a bar e ristoranti alle 22 o alle 23. Anche l’esercito sarà governo studia se rendere alcune misure più stringenti. impiegato nei controlli.
In particolare, saranno potenziati i controlli nei luoghi della movida o dove è più facile che si vadano a costituire afollamenti. Le operazioni di vigilanza saranno afidate C.3. XFORMAL-IT (F→I) anche ai militari impegnati nel progetto “Strade secure”.
Il ministro della Salute, Roberto Speranza, si sarebbe Input text: detto favorevole all’estensione dell’obbligo di mascherina Buongiorno, quella su E-Bay è molto bella. all’aperto a tutto il Paese, mantenendosi per quella linea di prudenza sostenuta fin dallo scoppio dell’emergenza. Camoscio generated text: La stessa considerazione, d’altronde, è condivisa anche Ciao, quella su E-Bay è davvero bella! dalla autorità locali, dal momento che diversi governatori hanno già emesso delle ordinanze che impongono di mT5-base generated text: coprire naso e bocca all’aperto, con multe salate per Buongiorno, quella su E-Bay è molto bella. C.4. XFORMAL-IT (I→F)
La crisi petrolifera del 1973 iniziò nell’ ottobre 1973 quando i membri dell’ Organizzazione dei Paesi esportatori di petrolio arabo (OAPEC, composta dai membri arabi dell’ OPEC più Egitto e Siria) proclamarono un embargo petrolifero. Alla fine dell’ embargo, nel marzo 1974, il prezzo del petrolio era salito da 3 dollari al barile a quasi 12 dollari a livello mondiale; i prezzi americani erano notevolmente più elevati. L’ embargo ha causato una crisi petrolifera, o “shock”, con molti efetti a breve e lungo termine sulla politica globale e sull’ economia globale. Più tardi fu chiamato il “primo shock petrolifero”, seguito dalla crisi petrolifera del 1979, definita il “secondo shock petrolifero”. Domanda: Quando è stata la seconda crisi petrolifera?