Andrés Marafioti

Hugo Laurençon, Andrés Marafioti, Victor Sanh et al.

The field of vision-language models (VLMs), which take images and texts as inputs and output texts, is rapidly evolving and has yet to reach consensus on several key aspects of the development pipeline, including data, architecture, and training methods. This paper can be seen as a tutorial for building a VLM. We begin by providing a comprehensive overview of the current state-of-the-art approaches, highlighting the strengths and weaknesses of each, addressing the major challenges in the field, and suggesting promising research directions for underexplored areas. We then walk through the practical steps to build Idefics3-8B, a powerful VLM that significantly outperforms its predecessor Idefics2-8B, while being trained efficiently, exclusively on open datasets, and using a straightforward pipeline. These steps include the creation of Docmatix, a dataset for improving document understanding capabilities, which is 240 times larger than previously available datasets. We release the model along with the datasets created for its training.

Loubna Ben Allal, Anton Lozhkov, Elie Bakouch et al. · huggingface

While large language models have facilitated breakthroughs in many applications of artificial intelligence, their inherent largeness makes them computationally expensive and challenging to deploy in resource-constrained settings. In this paper, we document the development of SmolLM2, a state-of-the-art "small" (1.7 billion parameter) language model (LM). To attain strong performance, we overtrain SmolLM2 on ~11 trillion tokens of data using a multi-stage training process that mixes web text with specialized math, code, and instruction-following data. We additionally introduce new specialized datasets (FineMath, Stack-Edu, and SmolTalk) at stages where we found existing datasets to be problematically small or low-quality. To inform our design decisions, we perform both small-scale ablations as well as a manual refinement process that updates the dataset mixing rates at each stage based on the performance at the previous stage. Ultimately, we demonstrate that SmolLM2 outperforms other recent small LMs including Qwen2.5-1.5B and Llama3.2-1B. To facilitate future research on LM development as well as applications of small LMs, we release both SmolLM2 as well as all of the datasets we prepared in the course of this project.

Andrés Marafioti, Orr Zohar, Miquel Farré et al.

Large Vision-Language Models (VLMs) deliver exceptional performance but require significant computational resources, limiting their deployment on mobile and edge devices. Smaller VLMs typically mirror design choices of larger models, such as extensive image tokenization, leading to inefficient GPU memory usage and constrained practicality for on-device applications. We introduce SmolVLM, a series of compact multimodal models specifically engineered for resource-efficient inference. We systematically explore architectural configurations, tokenization strategies, and data curation optimized for low computational overhead. Through this, we identify key design choices that yield substantial performance gains on image and video tasks with minimal memory footprints. Our smallest model, SmolVLM-256M, uses less than 1GB GPU memory during inference and outperforms the 300-times larger Idefics-80B model, despite an 18-month development gap. Our largest model, at 2.2B parameters, rivals state-of-the-art VLMs consuming twice the GPU memory. SmolVLM models extend beyond static images, demonstrating robust video comprehension capabilities. Our results emphasize that strategic architectural optimizations, aggressive yet efficient tokenization, and carefully curated training data significantly enhance multimodal performance, facilitating practical, energy-efficient deployments at significantly smaller scales.

Luis Wiedmann, Orr Zohar, Amir Mahla et al.

The advancement of vision-language models (VLMs) is hampered by a fragmented landscape of inconsistent and contaminated public datasets. We introduce FineVision, a meticulously collected, curated, and unified corpus of 24 million samples - the largest open resource of its kind. We unify more than 200 sources into 185 subsets via a semi-automated, human-in-the-loop pipeline: automation performs bulk ingestion and schema mapping, while reviewers audit mappings and spot-check outputs to verify faithful consumption of annotations, appropriate formatting and diversity, and safety; issues trigger targeted fixes and re-runs. The workflow further applies rigorous de-duplication within and across sources and decontamination against 66 public benchmarks. FineVision also encompasses agentic/GUI tasks with a unified action space; reviewers validate schemas and inspect a sample of trajectories to confirm executable fidelity. Models trained on FineVision consistently outperform those trained on existing open mixtures across a broad evaluation suite, underscoring the benefits of scale, data hygiene, and balanced automation with human oversight. We release the corpus and curation tools to accelerate data-centric VLM research.

Andrés Marafioti, Michel Hayoz, Mathias Gallardo et al.

Cataract surgery is a sight saving surgery that is performed over 10 million times each year around the world. With such a large demand, the ability to organize surgical wards and operating rooms efficiently is critical to delivery this therapy in routine clinical care. In this context, estimating the remaining surgical duration (RSD) during procedures is one way to help streamline patient throughput and workflows. To this end, we propose CataNet, a method for cataract surgeries that predicts in real time the RSD jointly with two influential elements: the surgeon's experience, and the current phase of the surgery. We compare CataNet to state-of-the-art RSD estimation methods, showing that it outperforms them even when phase and experience are not considered. We investigate this improvement and show that a significant contributor is the way we integrate the elapsed time into CataNet's feature extractor.

Andrés Marafioti, Nicki Holighaus, Piotr Majdak

In audio processing applications, phase retrieval (PR) is often performed from the magnitude of short-time Fourier transform (STFT) coefficients. Although PR performance has been observed to depend on the considered STFT parameters and audio data, the extent of this dependence has not been systematically evaluated yet. To address this, we studied the performance of three PR algorithms for various types of audio content and various STFT parameters such as redundancy, time-frequency ratio, and the type of window. The quality of PR was studied in terms of objective difference grade and signal-to-noise ratio of the STFT magnitude, to provide auditory- and signal-based quality assessments. Our results show that PR quality improved with increasing redundancy, with a strong relevance of the time-frequency ratio. The effect of the audio content was smaller but still observable. The effect of the window was only significant for one of the PR algorithms. Interestingly, for a good PR quality, each of the three algorithms required a different set of parameters, demonstrating the relevance of individual parameter sets for a fair comparison across PR algorithms. Based on these results, we developed guidelines for optimizing STFT parameters for a given application.

Andrés Marafioti, Nicki Holighaus, Nathanaël Perraudin et al.

Time-frequency (TF) representations provide powerful and intuitive features for the analysis of time series such as audio. But still, generative modeling of audio in the TF domain is a subtle matter. Consequently, neural audio synthesis widely relies on directly modeling the waveform and previous attempts at unconditionally synthesizing audio from neurally generated invertible TF features still struggle to produce audio at satisfying quality. In this article, focusing on the short-time Fourier transform, we discuss the challenges that arise in audio synthesis based on generated invertible TF features and how to overcome them. We demonstrate the potential of deliberate generative TF modeling by training a generative adversarial network (GAN) on short-time Fourier features. We show that by applying our guidelines, our TF-based network was able to outperform a state-of-the-art GAN generating waveforms directly, despite the similar architecture in the two networks.

Andrés Marafioti, Nicki Holighaus, Piotr Majdak et al.

We studied the ability of deep neural networks (DNNs) to restore missing audio content based on its context, a process usually referred to as audio inpainting. We focused on gaps in the range of tens of milliseconds. The proposed DNN structure was trained on audio signals containing music and musical instruments, separately, with 64-ms long gaps. The input to the DNN was the context, i.e., the signal surrounding the gap, transformed into time-frequency (TF) coefficients. Our results were compared to those obtained from a reference method based on linear predictive coding (LPC). For music, our DNN significantly outperformed the reference method, demonstrating a generally good usability of the proposed DNN structure for inpainting complex audio signals like music.