arXiv subjects

João Alves

Publications and source records attributed to João Alves.

At least 37 records · Page 2Linked to original sources

arXiv2025

EuroLLM-9B: Technical Report

This report presents EuroLLM-9B, a large language model trained from scratch to support the needs of European citizens by covering all 24 official European Union languages and 11 additional languages. EuroLLM addresses the issue of European languages being underrepresented and underserved in existing open large language models. We provide a comprehensive overview of EuroLLM-9B's development, including tokenizer design, architectural specifications, data filtering, and training procedures. We describe the pre-training data collection and filtering pipeline, including the creation of EuroFilter, an AI-based multilingual filter, as well as the design of EuroBlocks-Synthetic, a novel synthetic dataset for post-training that enhances language coverage for European languages. Evaluation results demonstrate EuroLLM-9B's competitive performance on multilingual benchmarks and machine translation tasks, establishing it as the leading open European-made LLM of its size. To support open research and adoption, we release all major components of this work, including the base and instruction-tuned models, the EuroFilter classifier, and the synthetic post-training dataset.

cs.CL↗

arXiv2025

Read My Ears! Horse Ear Movement Detection for Equine Affective State Assessment

The Equine Facial Action Coding System (EquiFACS) enables the systematic annotation of facial movements through distinct Action Units (AUs). It serves as a crucial tool for assessing affective states in horses by identifying subtle facial expressions associated with discomfort. However, the field of horse affective state assessment is constrained by the scarcity of annotated data, as manually labelling facial AUs is both time-consuming and costly. To address this challenge, automated annotation systems are essential for leveraging existing datasets and improving affective states detection tools. In this work, we study different methods for specific ear AU detection and localization from horse videos. We leverage past works on deep learning-based video feature extraction combined with recurrent neural networks for the video classification task, as well as a classic optical flow based approach. We achieve 87.5% classification accuracy of ear movement presence on a public horse video dataset, demonstrating the potential of our approach. We discuss future directions to develop these systems, with the aim of bridging the gap between automated AU detection and practical applications in equine welfare and veterinary diagnostics. Our code will be made publicly available at https://github.com/jmalves5/read-my-ears.

cs.CV↗

arXiv2025

Origin of the IRAS Vela Shell: New Insights from 3D Dust Mapping

The IRAS Vela Shell (IVS) is a structure of enhanced FIR emission located towards the Gum Nebula, a prominent region of $\rm Hα$ emission in the local Milky Way shaped by various galactic stellar feedback over the past several million years. We constrain the 3D spatial geometry of the IVS using a parsec-resolution 3D dust map and contextualize it within the broader Gum Nebula. Our analysis reveals a dense, bowl-like IVS structure below the Galactic plane, with a more diffuse component above. We obtain a total shell mass of $5.1_{-2.4}^{+2.4}\times 10^{4}\;\rm M_{\odot}$ and, incorporating previous studies on shell expansion, a momentum of $6.0_{-3.4}^{+4.7}\times 10^{5}\;\rm M_{\odot}\;km\; s^{-1}$. We find a spatial correlation between the morphology of the dust-traced IVS and the Gum Nebula's $\rm Hα$ emission when projected onto the sky. We quantify contributions of feedback from stellar winds, an expanding HII region, and supernovae to the IVS formation, finding that stellar winds are subdominant. Our momentum analysis shows that both an HII region and supernova feedback could drive the shell's expansion. Using astrometric constraints from Gaia and Hipparcos, we trace back nearby feedback sources and find that the massive stars $\gamma2$ Velorum and $ζ$ Puppis are currently within the IVS, producing enough ionizing luminosity to form an HII region of comparable size. Alternatively, if the IVS' momentum is primarily driven by supernovae, $1-2$ events would be required. We also identify several young massive clusters that could have hosted supernovae within the past 3 Myr.

astro-ph.GA↗

arXiv2025

CroissantLLM: A Truly Bilingual French-English Language Model

We introduce CroissantLLM, a 1.3B language model pretrained on a set of 3T English and French tokens, to bring to the research and industrial community a high-performance, fully open-sourced bilingual model that runs swiftly on consumer-grade local hardware. To that end, we pioneer the approach of training an intrinsically bilingual model with a 1:1 English-to-French pretraining data ratio, a custom tokenizer, and bilingual finetuning datasets. We release the training dataset, notably containing a French split with manually curated, high-quality, and varied data sources. To assess performance outside of English, we craft a novel benchmark, FrenchBench, consisting of an array of classification and generation tasks, covering various orthogonal aspects of model performance in the French Language. Additionally, rooted in transparency and to foster further Large Language Model research, we release codebases, and dozens of checkpoints across various model sizes, training data distributions, and training steps, as well as fine-tuned Chat models, and strong translation models. We evaluate our model through the FMTI framework, and validate 81 % of the transparency criteria, far beyond the scores of even most open initiatives. This work enriches the NLP landscape, breaking away from previous English-centric work in order to strengthen our understanding of multilinguality in language models.

cs.CL↗

arXiv2025

From moving groups to star formation in the Solar Neighborhood

Moving groups in the solar neighborhood are ensembles of co-moving stars, likely originating due to forces from spiral arms, the Galactic bar, or external perturbations. Their co-movement with young clusters indicates recent star formation within these moving groups, but a lack of precise three-dimensional position and velocity measurements has obscured this connection. Using backward orbit integrations of 509 clusters within 1 kpc - based on Gaia DR3 and supplemented with APOGEE-2 and GALAH DR3 radial velocities - we trace their evolution over the past 100 Myr. We find that most clusters separate into three spatial groups that each trace one of the Pleiades, Coma Berenices, and Sirius moving groups. The same trend is not seen for the Hyades moving group. The young clusters of the Alpha Persei, Messier 6, and Collinder 135 families of clusters, previously found to have formed in three massive star-forming complexes, co-move with either the Pleiades (Alpha Persei and Messier 6) or Coma Berenices (Collinder 135). Our results provide a sharper view of how large-scale Galactic dynamics have shaped recent, nearby star formation.

astro-ph.GA↗

arXiv2025

HP2 Survey V. Ophiuchus: Filament formation in a dispersing cloud complex

We search for potential ``birthmarks'' left from the formation of filamentary molecular clouds in the Ophiuchus complex. We use high dynamic-range column density and temperature maps derived from \textit{Herschel}, \textit{Planck}, and \textit{2MASS/NICEST} extinction data. We find two distinct types of filaments based on their orientation relative to nearby massive stars: radial (R-type) and tangential (T-type). R-type filaments exhibit decreasing mass profiles away from massive stars, while T-type filaments show flat but structured profiles. We propose a scenario where both filament types originate from the dynamic interplay of compression and stretching forces exerted by a fast outflow emanating from the OB association. The two formation mechanisms leave distinct observable ``birthmarks'' (namely, filament orientation, mass distribution, and star formation location) on each filament type. Our results illustrate a complex phase in molecular cloud evolution with two simultaneous yet contrasting processes: the formation of filaments and stars via the dispersal of residual gas from a previous massive star formation event. Our approach highlights the importance of taking into account the wider context of a star-forming complex, rather than concentrating exclusively on particular subregions.

astro-ph.GA↗

arXiv2024

The physical properties of Cluster Chains

We explore the kinematics and star formation history of the Scorpius Centaurus (Sco-Cen) OB association following the initial identification of sequential, linearly aligned chains of clusters. Building upon our characterization of the Corona Australis (CrA) chain, we now analyze two additional major cluster chains that exhibit similar characteristics: the Lower Centaurus Crux (LCC) and Upper Scorpius (Upper Sco) chains. All three cluster chains display distinct sequential patterns in 1) the 3D spatial distribution, 2) age, 3) velocity, and 4) mass. The Upper-Sco chain is the most massive and complex cluster chain, possibly consisting of two or more overlapping subchains. We discuss the possible formation of cluster chains and argue for a scenario where feedback from the most massive star formation episode 15 Myr ago initiated the formation of these spatio-temporal cluster sequences. Our results identify cluster chains as a distinct type of stellar structure with well-defined physical properties, formed in environments capable of sustaining stellar feedback over timescales of 5-10 Myr. We find that around 40% of the stellar population in Sco-Cen formed due to triggered star formation, with 35% forming along the three cluster chains. We conclude that cluster chains could be common structures in OB associations, particularly in regions that have similar natal environments as Sco-Cen. Beyond their significance for star formation and stellar feedback, they appear to be promising laboratories for chemical enrichment and the transport of elements from one generation to the next in the same star-forming region.

astro-ph.GA↗

arXiv2024

EuroLLM: Multilingual Language Models for Europe

The quality of open-weight LLMs has seen significant improvement, yet they remain predominantly focused on English. In this paper, we introduce the EuroLLM project, aimed at developing a suite of open-weight multilingual LLMs capable of understanding and generating text in all official European Union languages, as well as several additional relevant languages. We outline the progress made to date, detailing our data collection and filtering process, the development of scaling laws, the creation of our multilingual tokenizer, and the data mix and modeling configurations. Additionally, we release our initial models: EuroLLM-1.7B and EuroLLM-1.7B-Instruct and report their performance on multilingual general benchmarks and machine translation.

cs.CL↗

arXiv2024

The "C": The large Chameleon-Musca-Coalsack cloud

Recent advancements in 3D dust mapping have transformed our understanding of the Milky Way's local interstellar medium, enabling us to explore its structure in three spatial dimensions for the first time. In this Letter, we use the most recent 3D dust map by Edenhofer et al. to study the well-known Chameleon, Musca, and Coalsack cloud complexes, located about 200 pc from the Sun. We find that these three complexes are not isolated but rather connect to form a surprisingly well-defined half-ring, constituting a single C-shaped cloud with a radius of about 50 pc, a thickness of about 45 pc, and a total mass of about $5 \times 10^{4} \mathrm{M}_{\odot}$, or $9 \times 10^{4} \mathrm{M}_{\odot}$ if including everything in the vicinity of the C-shaped cloud. Despite the absence of an evident feedback source at its center, the dynamics of young stellar clusters associated with the C structure suggest that a single supernova explosion about 4 Myr to 10 Myr ago likely shaped this structure. Our findings support a single origin story for these cloud complexes, suggesting that they were formed by feedback-driven gas compression, and offer new insights into the processes that govern the birth of star-forming clouds in feedback-dominated regions, such as the Scorpius-Centaurus association.

astro-ph.GA↗

arXiv2024

Most nearby young star clusters formed in three massive complexes

Efforts to unveil the structure of the local interstellar medium and its recent star formation history have spanned the past seventy years. Recent studies utilizing precise data from space astrometry missions have revealed nearby, newly formed star clusters with connected origins. Nonetheless, mapping young clusters across the entire sky back to their natal regions has been hindered by a lack of clusters with precise radial velocity data. Here we show that 155 out of 272 (57 percent) high-quality young clusters within one kiloparsec of the Sun arise from three distinct spatial volumes. This conclusion is based upon the analysis of data from the third Gaia release and other large-scale spectroscopic surveys. Currently dispersed throughout the Solar Neighborhood, their past positions over 30 Myr ago reveal that these families of clusters each formed in one of three compact, massive star-forming complexes. One of these families includes all of the young clusters near the Sun -- the Taurus and Sco-Cen star-forming complexes. We estimate that over 200 supernovae were produced from these families and argue that these clustered supernovae produced both the Local Bubble and the largest nearby supershell GSH 238+00+09, both of which are clearly visible in modern three-dimensional dust maps.

astro-ph.GA↗

arXiv2024

Isoleucine gate blocks K+ conduction in C-type inactivation

Many voltage-gated potassium (Kv) channels display a time-dependent phenomenon called C-type inactivation, whereby prolonged activation by voltage leads to the inhibition of ionic conduction, a process that involves a conformational change at the selectivity filter toward a non-conductive state. Recently, a high-resolution structure of a strongly inactivating triple-mutant channel kv1.2-kv2.1-3m revealed a novel conformation of the selectivity filter that is dilated at its outer end, distinct from the well-characterized conductive state. While the experimental structure was interpreted as the elusive non-conductive state, molecular dynamics simulations and electrophysiology measurements demonstrate that the dilated filter of kv1.2-kv2.1-3m, however, is conductive and, as such, cannot completely account for the inactivation of the channel observed in functional experiments. An additional conformational change implicating isoleucine residues at position 398 along the pore lining segment S6 is required to effectively block ion conduction. It is shown that the I398 residues from the four subunits act as a state-dependent hydrophobic gate located immediately beneath the selectivity filter. As a critical piece of the C-type inactivation machinery, this structural feature is the potential target of a broad class of QA blockers and negatively charged activators thus opening new research directions towards the development of drugs that specifically modulate gating-states of Kv channels.

physics.bio-ph↗

arXiv2024

Tower: An Open Multilingual Large Language Model for Translation-Related Tasks

While general-purpose large language models (LLMs) demonstrate proficiency on multiple tasks within the domain of translation, approaches based on open LLMs are competitive only when specializing on a single task. In this paper, we propose a recipe for tailoring LLMs to multiple tasks present in translation workflows. We perform continued pretraining on a multilingual mixture of monolingual and parallel data, creating TowerBase, followed by finetuning on instructions relevant for translation processes, creating TowerInstruct. Our final model surpasses open alternatives on several tasks relevant to translation workflows and is competitive with general-purpose closed LLMs. To facilitate future research, we release the Tower models, our specialization dataset, an evaluation framework for LLMs focusing on the translation ecosystem, and a collection of model generations, including ours, on our benchmark.

cs.CL↗

arXiv2024

The Radcliffe Wave is Oscillating

Our Sun lies within 300 pc of the 2.7-kpc-long sinusoidal chain of dense gas clouds known as the Radcliffe Wave. The structure's wave-like shape was discovered using 3D dust mapping, but initial kinematic searches for oscillatory motion were inconclusive. Here we present evidence that the Radcliffe Wave is oscillating through the Galactic plane while also drifting radially away from the Galactic Center. We use measurements of line-of-sight velocity for 12CO and 3D velocities of young stellar clusters to show that the most massive star-forming regions spatially associated with the Radcliffe Wave (including Orion, Cepheus, North America, and Cygnus X) move as if they are part of an oscillating wave driven by the gravitational acceleration of the Galactic potential. By treating the Radcliffe Wave as a coherently oscillating structure, we can derive its motion independently of the local Galactic mass distribution, and directly measure local properties of the Galactic potential as well as the Sun's vertical oscillation period. In addition, the measured drift of the Radcliffe Wave radially outward from the Galactic Center suggests that the cluster whose supernovae ultimately created today's expanding Local Bubble may have been born in the Radcliffe Wave.

astro-ph.GA↗

arXiv2024

A Parsec-Scale Catalog of Molecular Clouds in the Solar Neighborhood Based on 3D Dust Mapping: Implications for the Mass-Size Relation

We dendrogram the Leike et al. 2020 3D dust map, leveraging its $\sim 1$ pc spatial resolution to produce a uniform catalog of molecular clouds in the solar neighborhood. Using accurate distances, we measure the properties of 65 clouds in true 3D space, eliminating much of the uncertainty in mass, size, and density. Clouds in the catalog contain a total of $1.1 \times 10^5 \; \rm M_{\odot}$, span distances of $116-440$ pc, and include a dozen well-studied clouds in the literature. In addition to deriving cloud properties in 3D volume density space, we create 2D dust extinction maps from the 3D data by projecting the 3D clouds onto a 2D "Sky" view. We measure the properties of the 2D clouds separately from the 3D clouds. We compare the scaling relation between the masses and sizes of clouds following Larson 1981. We find that our 2D projected mass-size relation, $M \propto r^{2.1}$, agrees with Larson's Third Relation, but our 3D derived properties lead to a scaling relation of about one order larger: $M \propto r^{2.9}$. Validating predictions from theory and numerical simulations, our results indicate that the mass-size relation is sensitive to whether column or volume density is used to define clouds, since mass scales with area in 2D ($M \propto r^{2}$) and with volume in 3D ($M \propto r^{3}$). Our results imply a roughly constant column and volume density in 2D and 3D, respectively, for molecular clouds, as would be expected for clouds where the lower density, larger volume-filling gas dominates the cloud mass budget.

astro-ph.GA↗

arXiv2023

Insights into star formation and dispersal from the synchronisation of stellar clocks

Age is one of the most fundamental parameters of stars, yet it is one of the hardest to determine as it requires modelling various aspects of stellar formation and evolution. When we compare the ages derived from isochronal and dynamical traceback methods for six young stellar associations, we find a systematic discrepancy. Specifically, dynamical traceback ages are consistently younger by an average of $\langleΔ_{\rm Age}\rangle = 5.5 \pm 1.1$ Myr. We rule out measurement errors as the cause of the age mismatch and propose that $Δ_{\rm Age}$ indicates the time a young star remains bound to its parental cloud before moving away from its siblings. In this framework, the dynamical traceback "clock" starts when a stellar cluster or association begins to expand after expelling most of the gas, while the isochronal "clock" starts earlier when most stars form. The difference between these two age-dating techniques is a powerful tool to constraint evolutionary models, as isochronal ages cannot be younger than dynamical traceback ages. Measuring the $Δ_{\rm Age}$ accurately and understanding its variations across different environments will provide further information on the impact of local conditions and stellar feedback on the formation and dispersal of stellar clusters.

astro-ph.SR↗

arXiv2023

Steering Large Language Models for Machine Translation with Finetuning and In-Context Learning

Large language models (LLMs) are a promising avenue for machine translation (MT). However, current LLM-based MT systems are brittle: their effectiveness highly depends on the choice of few-shot examples and they often require extra post-processing due to overgeneration. Alternatives such as finetuning on translation instructions are computationally expensive and may weaken in-context learning capabilities, due to overspecialization. In this paper, we provide a closer look at this problem. We start by showing that adapter-based finetuning with LoRA matches the performance of traditional finetuning while reducing the number of training parameters by a factor of 50. This method also outperforms few-shot prompting and eliminates the need for post-processing or in-context examples. However, we show that finetuning generally degrades few-shot performance, hindering adaptation capabilities. Finally, to obtain the best of both worlds, we propose a simple approach that incorporates few-shot examples during finetuning. Experiments on 10 language pairs show that our proposed approach recovers the original few-shot capabilities while keeping the added benefits of finetuning.

cs.CL↗

arXiv2023

The star formation history of the Sco-Cen association. Coherent star formation patterns in space and time

We reconstruct the star formation history of the Sco-Cen OB association using a novel high-resolution age map of the region. We develop an approach to produce robust ages for Sco-Cen's recently identified 37 stellar clusters using the SigMA algorithm. The Sco-Cen star formation timeline reveals four periods of enhanced star formation activity, or bursts, remarkably separated by about 5 Myr. Of these, the second burst, which occurred about 15 million years ago, is by far the dominant, and most of Sco-Cen's stars and clusters were in place by the end of this burst. The formation of stars and clusters in Sco-Cen is correlated but not linearly, implying that more stars were formed per cluster during the peak of the star formation rate. Most of the clusters that are large enough to have supernova precursors were formed during the 15 Myr period. Star and cluster formation activity has been continuously declining since then. We have clear evidence that Sco-Cen formed from the inside out and contains 100-pc long chains of contiguous clusters exhibiting well-defined age gradients, from massive older clusters to smaller young clusters. These observables suggest an important role for feedback in forming about half of Sco-Cen stars, although follow-up work is needed to quantify this statement. Finally, we confirm that the Upper-Sco age controversy discussed in the literature during the last decades is solved: the nine clusters previously lumped together as Upper-Sco, a benchmark region for planet formation studies, exhibit a wide range of ages from 3 to 19 Myr.

astro-ph.SR↗

arXiv2023

The Solar Neighborhood in the Age of Gaia

Most of what we know about the formation of stars, and essentially everything we know about the formation of planets, comes from observations of our solar neighborhood within 2 kpc of the Sun. Before 2018, accurate distance measurements needed to turn the 2D Sky into a faithful 3D physical picture of the distribution of stars, and the interstellar matter that forms them, were few and far between. Here, we offer a holistic review of how, since 2018, data from the Gaia mission are revealing previously unseen and often unexpected 3D distributions of gas, dust, and young stars in the solar neighborhood. We summarize how new extinction-based techniques yield 3D dust maps and how the density structure mapped out offers key context for measuring young stars' 3D positions from Gaia and VLBI. We discuss how a subset of young stars in Gaia with measured radial velocities and proper motions is being used to recover 3D cloud motion and characterize the internal dynamics of individual star-forming regions. We review relationships between newly-identified clusters and streams of young stars and the molecular interstellar medium from which they evolve. The combination of these measures of gas and stars' 3D distribution and 3D motions provides unprecedented data for comparison with simulations and reframes our understanding of local star formation in a larger Galactic context. This new 3D view of our solar neighborhood in the age of Gaia shows that star-forming regions once thought to be isolated are often connected on kiloparsec scales, causing us to reconsider models for the arrangement of gas and young stars in galaxies.

astro-ph.GA↗