Jordi Armengol-Estapé

FAIR CodeGen team, Jade Copet, Quentin Carbonneaux et al. · meta-ai

We release Code World Model (CWM), a 32-billion-parameter open-weights LLM, to advance research on code generation with world models. To improve code understanding beyond what can be learned from training on static code alone, we mid-train CWM on a large amount of observation-action trajectories from Python interpreter and agentic Docker environments, and perform extensive multi-task reasoning RL in verifiable coding, math, and multi-turn software engineering environments. With CWM, we provide a strong testbed for researchers to explore the opportunities world modeling affords for improving code generation with reasoning and planning in computational environments. We present first steps of how world models can benefit agentic coding, enable step-by-step simulation of Python code execution, and show early results of how reasoning can benefit from the latter. CWM is a dense, decoder-only LLM trained with a context size of up to 131k tokens. Independent of its world modeling capabilities, CWM offers strong performance on general coding and math tasks: it reaches pass@1 scores of 65.8% on SWE-bench Verified (with test-time scaling), 68.6% on LiveCodeBench, 96.6% on Math-500, and 76.0% on AIME 2024. To support further research on code world modeling, we release model checkpoints after mid-training, SFT, and RL.

Asier Gutiérrez-Fandiño, David Pérez-Fernández, Jordi Armengol-Estapé et al.

In the recent years, transformer-based models have lead to significant advances in language modelling for natural language processing. However, they require a vast amount of data to be (pre-)trained and there is a lack of corpora in languages other than English. Recently, several initiatives have presented multilingual datasets obtained from automatic web crawling. However, the results in Spanish present important shortcomings, as they are either too small in comparison with other languages, or present a low quality derived from sub-optimal cleaning and deduplication. In this paper, we introduce esCorpius, a Spanish crawling corpus obtained from near 1 Pb of Common Crawl data. It is the most extensive corpus in Spanish with this level of quality in the extraction, purification and deduplication of web textual content. Our data curation process involves a novel highly parallel cleaning pipeline and encompasses a series of deduplication mechanisms that together ensure the integrity of both document and paragraph boundaries. Additionally, we maintain both the source web page URL and the WARC shard origin URL in order to complain with EU regulations. esCorpius has been released under CC BY-NC-ND 4.0 license and is available on HuggingFace.

Casimiro Pio Carrino, Jordi Armengol-Estapé, Asier Gutiérrez-Fandiño et al.

This work presents biomedical and clinical language models for Spanish by experimenting with different pretraining choices, such as masking at word and subword level, varying the vocabulary size and testing with domain data, looking for better language representations. Interestingly, in the absence of enough clinical data to train a model from scratch, we applied mixed-domain pretraining and cross-domain transfer approaches to generate a performant bio-clinical model suitable for real-world clinical data. We evaluated our models on Named Entity Recognition (NER) tasks for biomedical documents and challenging hospital discharge reports. When compared against the competitive mBERT and BETO models, we outperform them in all NER tasks by a significant margin. Finally, we studied the impact of the model's vocabulary on the NER performances by offering an interesting vocabulary-centric analysis. The results confirm that domain-specific pretraining is fundamental to achieving higher performances in downstream NER tasks, even within a mid-resource scenario. To the best of our knowledge, we provide the first biomedical and clinical transformer-based pretrained language models for Spanish, intending to boost native Spanish NLP applications in biomedicine. Our best models are freely available in the HuggingFace hub: https://huggingface.co/BSC-TeMU.

Chris Cummins, Volker Seeker, Jordi Armengol-Estapé et al. · meta-ai

Tools for rewriting, refactoring and optimizing code should be fast and correct. Large language models (LLMs), by their nature, possess neither of these qualities. Yet, there remains tremendous opportunity in using LLMs to improve code. We explore the use of LLMs not to transform code, but to code transforms. We propose a chain-of-thought approach to synthesizing code transformations from a small number of input/output code examples that incorporates execution and feedback. Unlike the direct rewrite approach, LLM-generated transformations are easy to inspect, debug, and validate. The logic of the rewrite is explicitly coded and easy to adapt. The compute required to run code transformations is minute compared to that of LLM rewriting. We test our approach on 16 Python code transformations and find that LLM- generated transforms are perfectly precise for 7 of them and less imprecise than direct LLM rewriting on the others. We hope to encourage further research to improving the precision of LLM code rewriting.

Jordi Armengol-Estapé, Rodrigo C. O. Rocha, Jackson Woodruff et al.

The escalating demand to migrate legacy software across different Instruction Set Architectures (ISAs) has driven the development of assembly-to-assembly translators to map between their respective assembly languages. However, the development of these tools requires substantial engineering effort. State-of-the-art approaches use lifting, a technique where source assembly code is translated to an architecture-independent intermediate representation (IR) (for example, the LLVM IR) and use a pre-existing compiler to recompile the IR to the target ISA. However, the hand-written rules these lifters employ are sensitive to the particular compiler and optimization level used to generate the code and require significant engineering effort to support each new ISA. We propose Forklift, the first neural lifter that learns how to translate assembly to LLVM IR using a token-level encoder-decoder Transformer. We show how to incrementally add support to new ISAs by fine tuning the assembly encoder and freezing the IR decoder, improving the overall accuracy and efficiency. We collect millions of parallel LLVM IR, x86, ARM, and RISC-V programs across compilers and optimization levels to train Forklift and set up an input/output-based accuracy harness. We evaluate Forklift on two challenging benchmark suites and translate 2.5x more x86 programs than a state-of-the-art hand-written lifter and 4.4x more x86 programs than GPT-4 as well as enabling translation from new ISAs.

Jordi Armengol-Estapé, Quentin Carbonneaux, Tianjun Zhang et al. · meta-ai

Code generation and understanding are critical capabilities for large language models (LLMs). Thus, most LLMs are pretrained and fine-tuned on code data. However, these datasets typically treat code as static strings and rarely exploit the dynamic information about their execution. Building upon previous work on trace modeling, we study Execution Tuning (E.T.), a training procedure in which we explicitly model real-world program execution traces without requiring manual test annotations. We train and evaluate models on different execution trace granularities (line and instruction-level) and strategies on the task of output prediction, obtaining around 80% accuracy on CruxEval and MBPP, and showing the advantages of dynamic scratchpads (i.e., self-contained intermediate computations updated by the model rather than accumulated as a history of past computations) on long executions (up to 14k steps). Finally, we discuss E.T.'s practical applications.

Jordi Armengol-Estapé, Jackson Woodruff, Chris Cummins et al.

Decompilation is a well-studied area with numerous high-quality tools available. These are frequently used for security tasks and to port legacy code. However, they regularly generate difficult-to-read programs and require a large amount of engineering effort to support new programming languages and ISAs. Recent interest in neural approaches has produced portable tools that generate readable code. However, to-date such techniques are usually restricted to synthetic programs without optimization, and no models have evaluated their portability. Furthermore, while the code generated may be more readable, it is usually incorrect. This paper presents SLaDe, a Small Language model Decompiler based on a sequence-to-sequence transformer trained over real-world code. We develop a novel tokenizer and exploit no-dropout training to produce high-quality code. We utilize type-inference to generate programs that are more readable and accurate than standard analytic and recent neural approaches. Unlike standard approaches, SLaDe can infer out-of-context types and unlike neural approaches, it generates correct code. We evaluate SLaDe on over 4,000 functions from ExeBench on two ISAs and at two optimizations levels. SLaDe is up to 6 times more accurate than Ghidra, a state-of-the-art, industrial-strength decompiler and up to 4 times more accurate than the large language model ChatGPT and generates significantly more readable code than both.

Ona de Gibert, Ksenia Kharitonova, Blanca Calvo Figueras et al.

In this work, we introduce sequence-to-sequence language resources for Catalan, a moderately under-resourced language, towards two tasks, namely: Summarization and Machine Translation (MT). We present two new abstractive summarization datasets in the domain of newswire. We also introduce a parallel Catalan-English corpus, paired with three different brand new test sets. Finally, we evaluate the data presented with competing state of the art models, and we develop baselines for these tasks using a newly created Catalan BART. We release the resulting resources of this work under open license to encourage the development of language technology in Catalan.

Asier Gutiérrez-Fandiño, David Pérez-Fernández, Jordi Armengol-Estapé

In this work, we present the Large Labelled Logo Dataset (L3D), a multipurpose, hand-labelled, continuously growing dataset. It is composed of around 770k of color 256x256 RGB images extracted from the European Union Intellectual Property Office (EUIPO) open registry. Each of them is associated to multiple labels that classify the figurative and textual elements that appear in the images. These annotations have been classified by the EUIPO evaluators using the Vienna classification, a hierarchical classification of figurative marks. We suggest two direct applications of this dataset, namely, logo classification and logo generation.

Asier Gutiérrez-Fandiño, Miquel Noguer i Alonso, Petter Kolm et al.

We introduce a new language representation model in finance called Financial Embedding Analysis of Sentiment (FinEAS). In financial markets, news and investor sentiment are significant drivers of security prices. Thus, leveraging the capabilities of modern NLP approaches for financial sentiment analysis is a crucial component in identifying patterns and trends that are useful for market participants and regulators. In recent years, methods that use transfer learning from large Transformer-based language models like BERT, have achieved state-of-the-art results in text classification tasks, including sentiment analysis using labelled datasets. Researchers have quickly adopted these approaches to financial texts, but best practices in this domain are not well-established. In this work, we propose a new model for financial sentiment analysis based on supervised fine-tuned sentence embeddings from a standard BERT model. We demonstrate our approach achieves significant improvements in comparison to vanilla BERT, LSTM, and FinBERT, a financial domain specific BERT.

Asier Gutiérrez-Fandiño, Jordi Armengol-Estapé, Aitor Gonzalez-Agirre et al.

There are many Language Models for the English language according to its worldwide relevance. However, for the Spanish language, even if it is a widely spoken language, there are very few Spanish Language Models which result to be small and too general. Legal slang could be think of a Spanish variant on its own as it is very complicated in vocabulary, semantics and phrase understanding. For this work we gathered legal-domain corpora from different sources, generated a model and evaluated against Spanish general domain tasks. The model provides reasonable results in those tasks.

Casimiro Pio Carrino, Jordi Armengol-Estapé, Ona de Gibert Bonet et al.

We introduce CoWeSe (the Corpus Web Salud Español), the largest Spanish biomedical corpus to date, consisting of 4.5GB (about 750M tokens) of clean plain text. CoWeSe is the result of a massive crawler on 3000 Spanish domains executed in 2020. The corpus is openly available and already preprocessed. CoWeSe is an important resource for biomedical and health NLP in Spanish and has already been employed to train domain-specific language models and to produce word embbedings. We released the CoWeSe corpus under a Creative Commons Attribution 4.0 International license, both in Zenodo (\url{https://zenodo.org/record/4561971\#.YTI5SnVKiEA}).

Jordi Armengol-Estapé, Ona de Gibert Bonet, Maite Melero

Generative Pre-trained Transformers (GPTs) have recently been scaled to unprecedented sizes in the history of machine learning. These models, solely trained on the language modeling objective, have been shown to exhibit outstanding few-shot learning capabilities in a number of different tasks. Nevertheless, aside from anecdotal experiences, little is known regarding their multilingual capabilities, given the fact that the pre-training corpus is almost entirely composed of English text. In this work, we investigate the multilingual skills of GPT-3, focusing on one language that barely appears in the pre-training corpus, Catalan, which makes the results especially meaningful; we assume that our results may be relevant for other languages as well. We find that the model shows an outstanding performance, particularly in generative tasks, with predictable limitations mostly in language understanding tasks but still with remarkable results given the zero-shot scenario. We investigate its potential and limits in extractive question-answering and natural language generation, as well as the effect of scale in terms of model size.

Jordi Armengol-Estapé, Michael F. P. O'Boyle

Deep learning has had a significant impact on many fields. Recently, code-to-code neural models have been used in code translation, code refinement and decompilation. However, the question of whether these models can automate compilation has yet to be investigated. In this work, we explore neural compilation, building and evaluating Transformer models that learn how to produce x86 assembler from C code. Although preliminary results are relatively weak, we make our data, models and code publicly available to encourage further research in this area.

Jordi Armengol-Estapé, Casimiro Pio Carrino, Carlos Rodriguez-Penagos et al.

Multilingual language models have been a crucial breakthrough as they considerably reduce the need of data for under-resourced languages. Nevertheless, the superiority of language-specific models has already been proven for languages having access to large amounts of data. In this work, we focus on Catalan with the aim to explore to what extent a medium-sized monolingual language model is competitive with state-of-the-art large multilingual models. For this, we: (1) build a clean, high-quality textual Catalan corpus (CaText), the largest to date (but only a fraction of the usual size of the previous work in monolingual language models), (2) train a Transformer-based language model for Catalan (BERTa), and (3) devise a thorough evaluation in a diversity of settings, comprising a complete array of downstream tasks, namely, Part of Speech Tagging, Named Entity Recognition and Classification, Text Classification, Question Answering, and Semantic Textual Similarity, with most of the corresponding datasets being created ex novo. The result is a new benchmark, the Catalan Language Understanding Benchmark (CLUB), which we publish as an open resource, together with the clean textual corpus, the language model, and the cleaning pipeline. Using state-of-the-art multilingual models and a monolingual model trained only on Wikipedia as baselines, we consistently observe the superiority of our model across tasks and settings.

Asier Gutiérrez-Fandiño, Jordi Armengol-Estapé, Marc Pàmies et al.

This work presents MarIA, a family of Spanish language models and associated resources made available to the industry and the research community. Currently, MarIA includes RoBERTa-base, RoBERTa-large, GPT2 and GPT2-large Spanish language models, which can arguably be presented as the largest and most proficient language models in Spanish. The models were pretrained using a massive corpus of 570GB of clean and deduplicated texts with 135 billion words extracted from the Spanish Web Archive crawled by the National Library of Spain between 2009 and 2019. We assessed the performance of the models with nine existing evaluation datasets and with a novel extractive Question Answering dataset created ex novo. Overall, MarIA models outperform the existing Spanish models across a variety of NLU tasks and training settings.

Asier Gutiérrez-Fandiño, David Pérez-Fernández, Jordi Armengol-Estapé et al.

The training of neural networks is usually monitored with a validation (holdout) set to estimate the generalization of the model. This is done instead of measuring intrinsic properties of the model to determine whether it is learning appropriately. In this work, we suggest studying the training of neural networks with Algebraic Topology, specifically Persistent Homology (PH). Using simplicial complex representations of neural networks, we study the PH diagram distance evolution on the neural network learning process with different architectures and several datasets. Results show that the PH diagram distance between consecutive neural network states correlates with the validation accuracy, implying that the generalization error of a neural network could be intrinsically estimated without any holdout set.

Asier Gutiérrez-Fandiño, Jordi Armengol-Estapé, Casimiro Pio Carrino et al.

We computed both Word and Sub-word Embeddings using FastText. For Sub-word embeddings we selected Byte Pair Encoding (BPE) algorithm to represent the sub-words. We evaluated the Biomedical Word Embeddings obtaining better results than previous versions showing the implication that with more data, we obtain better representations.

David Pérez-Fernández, Asier Gutiérrez-Fandiño, Jordi Armengol-Estapé et al.

Characterizing the structural properties of neural networks is crucial yet poorly understood, and there are no well-established similarity measures between networks. In this work, we observe that neural networks can be represented as abstract simplicial complex and analyzed using their topological 'fingerprints' via Persistent Homology (PH). We then describe a PH-based representation proposed for characterizing and measuring similarity of neural networks. We empirically show the effectiveness of this representation as a descriptor of different architectures in several datasets. This approach based on Topological Data Analysis is a step towards better understanding neural networks and serves as a useful similarity measure.

Asier Gutiérrez-Fandiño, Jordi Armengol-Estapé, Marta Villegas

Researchers that work for the same institution use their email as the main communication tool. Email can be one of the most fruitful attack vectors of research institutions as they also contain access to all accounts and thus to all private information. We propose an approach for analyzing in terms of security research institutions' communication networks. We first obtained institutions' communication networks as well as a method to analyze possible breaches of collected emails. We downloaded the network of 4 different research centers, three from Spain and one from Portugal. We then ran simulations of Susceptible-Exposed-Infected-Recovered (SEIR) complex network dynamics model for analyzing the vulnerability of the network. More than half of the nodes have more than one security breach, and our simulation results show that more than 90\% of the networks' nodes are vulnerable. This method can be employed for enhancing security of research centers and can make email accounts' use security-aware. It may additionally open new research lines in communication security. Finally, we manifest that, due to confidentiality reasons, the sources we utilized for obtaining communication networks should not be providing the information that we were able to gather.

Jordi Armengol-Estapé, Marta R. Costa-jussà, Carlos Escolano

Introducing factors, that is to say, word features such as linguistic information referring to the source tokens, is known to improve the results of neural machine translation systems in certain settings, typically in recurrent architectures. This study proposes enhancing the current state-of-the-art neural machine translation architecture, the Transformer, so that it allows to introduce external knowledge. In particular, our proposed modification, the Factored Transformer, uses linguistic factors that insert additional knowledge into the machine translation system. Apart from using different kinds of features, we study the effect of different architectural configurations. Specifically, we analyze the performance of combining words and features at the embedding level or at the encoder level, and we experiment with two different combination strategies. With the best-found configuration, we show improvements of 0.8 BLEU over the baseline Transformer in the IWSLT German-to-English task. Moreover, we experiment with the more challenging FLoRes English-to-Nepali benchmark, which includes both extremely low-resourced and very distant languages, and obtain an improvement of 1.2 BLEU.