arXiv subjects

Maximilian Böther

Publications and source records attributed to Maximilian Böther.

15 recordsLinked to original sources

arXiv2026

Zephon: Elastic Determinism for Online, Stateful Foundation Model Data Loading Pipelines

Deterministic data loading is important for foundation model development: model researchers need confidence that differences they observe across costly ablations are caused by the parameter they changed rather than non-determinism in the training data sequence. The data loader must provide elastic determinism, i.e., a deterministic sequence of global training data batches despite changes to the GPU topology across runs (e.g., due to GPU scarcity), frequent checkpoint-resume cycles, and different data processing execution backends. Achieving this is difficult because modern foundation model data pipelines tokenize, pack, and mix samples online, introducing stateful n-to-m transformations that break sample indexing. Existing data loaders largely assume indexable 1-to-1 pipelines, and the common workaround of offline materialization is expensive and, for some modalities such as video, infeasible. We present Zephon, a data loader for foundation models that supports online, stateful pipelines while providing elastic determinism and efficient resumption from checkpoints. It partitions the global stream into topology-independent lanes, serializes ordering decisions while parallelizing stateless work on interchangeable backends, and checkpoints only bounded in-flight state so recovery cost does not grow with training progress. We evaluate Zephon on text and vision-language workloads and show that it achieves competitive throughput while providing a combination of guarantees that no existing loader offers for online, stateful pipelines.

cs.LG↗

arXiv2026

DataComp-VLM: Improved Open Datasets for Vision-Language Models

Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies. We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training. As part of DCVLM, we collect 160 datasets spanning four data types -- image-caption pairs, multimodal interleaved documents, text-only, and instruction-tuning data -- into a corpus of 6T multimodal tokens. DCVLM allows participants to test curation strategies (filtering, mixing, formatting, sampling) across 1B-8B models and 6.25B-200B token budgets. Models are then evaluated on a carefully selected suite of up to 52 downstream benchmarks across 9 domains. We conduct extensive experiments on DCVLM and find that data mixing, not filtering, is key to a high-quality training dataset: instruction-heavy mixtures scale better than caption-heavy ones, with gains widening at larger scales. The resulting dataset, DCVLM-Baseline, enables training an 8B VLM to 63.6% accuracy on our 33-task core suite with 200B training tokens. Compared to FineVision, the state-of-the-art open VLM training dataset, this represents an improvement of +5.4pp. DCVLM and all accompanying artifacts will be made publicly available at https://www.datacomp.ai/dcvlm/.

cs.CV↗

arXiv2026

20/20 Vision Language Models: A Prescription for Better VLMs through Data Curation Alone

Data curation has shifted the quality-compute frontier for language-model and contrastive image-text pretraining, but its role for vision-language models (VLMs) is far less established. We ask how far data curation alone can take VLM performance, holding architecture, training recipe, and compute fixed and varying only the training data. Our pipeline, applied to the MAmmoTH-VL single-image subset, lifts performance by +11.7pp on average across 20 public VLM benchmarks (spanning grounding, VQA, OCR/documents, captioning, spatial/3D, counting, charts, math, brand-ID, and multi-image reasoning) and by +11.3pp on average across all nine capability axes of DatBench, our high-fidelity VLM eval suite. At 2B, our curated model surpasses InternVL3.5-2B by 9.9pp at ~17x less training compute and closes the gap to Qwen3-VL-2B to within 1.8pp at ~87x less compute, from pretraining alone. Beyond accuracy, curation delivers four further properties: (1) Reliability: per-capability std across training seeds drops by ~67% and the lift survives a 4k-to-16k context-length sweep; (2) OOD generalization: the 9-eval OOD average rises by +7.2pp, and multi-image BLINK rises by +3.09pp despite single-image-only training, with Visual Correspondence gaining +11.8pp; (3) Behavioral gains beyond benchmarks: across ~1,100 open-ended queries the curated 2B is more honest and more specific than the matched-compute baseline, and more concise and less refusal-prone than a frontier 2B reference; (4) Pareto-dominance on inference cost: at every scale (1B, 2B, 4B) the curated model raises accuracy while lowering response FLOPs vs. the matched-compute baseline, and the curated 4B matches near-frontier accuracy at 3.3x lower response FLOPs than Qwen3-VL-4B. Data curation is a high-leverage tool for building better VLMs, reaching near-frontier accuracy at up to ~150x less training compute.

cs.LG↗

arXiv2026

The Finetuner's Fallacy: When to Pretrain with Your Finetuning Data

Real-world model deployments demand strong performance on narrow domains where data is often scarce. Typically, practitioners finetune models to specialize them, but this risks overfitting to the domain and forgetting general knowledge. We study a simple strategy, specialized pretraining (SPT), where a small domain dataset, typically reserved for finetuning, is repeated starting from pretraining as a fraction of the total tokens. Across three specialized domains (ChemPile, MusicPile, and ProofPile), SPT improves domain performance and preserves general capabilities after finetuning compared to standard pretraining. In our experiments, SPT reduces the pretraining tokens needed to reach a given domain performance by up to 1.75x. These gains grow when the target domain is underrepresented in the pretraining corpus: on domains far from web text, a 1B SPT model outperforms a 3B standard pretrained model. Beyond these empirical gains, we derive overfitting scaling laws to guide practitioners in selecting the optimal domain-data repetition for a given pretraining compute budget. Our observations reveal the finetuner's fallacy: while finetuning may appear to be the cheapest path to domain adaptation, introducing specialized domain data during pretraining stretches its utility. SPT yields better specialized domain performance (via reduced overfitting across repeated exposures) and better general domain performance (via reduced forgetting during finetuning), ultimately achieving stronger results with fewer parameters and less total compute when amortized over inference. To get the most out of domain data, incorporate it as early in training as possible.

cs.LG↗

arXiv2026

ÜberWeb: Insights from Multilingual Curation for a 20-Trillion-Token Dataset

Multilinguality is a core capability for modern foundation models, yet training high-quality multilingual models remains challenging due to uneven data availability across languages. A further challenge is the performance interference that can arise from joint multilingual training, commonly referred to as the "curse of multilinguality". We study multilingual data curation across thirteen languages and find that many reported regressions are not inherent to multilingual scaling but instead stem from correctable deficiencies in data quality and composition rather than fundamental capacity limits. In controlled bilingual experiments, improving data quality for any single language benefits others: curating English improves non-English performance in 12 of 13 languages, while curating non-English yields reciprocal improvements in English. Bespoke per-language curation produces substantially larger within-language improvements. Extending these findings to large-scale general-purpose training mixtures, we show that curated multilingual allocations comprising under 8% of total tokens remain remarkably effective. We operationalize this approach within an effort that produced a 20T-token pretraining corpus derived entirely from public sources. Models with 3B and 8B parameters trained on a 1T-token random subset achieve competitive multilingual accuracy with 4-10x fewer training FLOPs than strong public baselines, establishing a new Pareto frontier in multilingual performance versus compute. Moreover, these benefits extend to frontier model scale: the 20T-token corpus served as part of the pretraining dataset for Trinity Large (400B/A13B), which exhibits strong multilingual performance relative to its training FLOPs. These results show that targeted, per-language data curation mitigates multilingual interference and enables compute-efficient multilingual scaling.

cs.LG↗

arXiv2026

Mixtera: A Data Plane for Foundation Model Training

State-of-the-art large language and vision models are trained over trillions of tokens that are aggregated from a large variety of sources. As training data collections grow, manually managing the samples becomes time-consuming, tedious, and prone to errors. Yet recent research shows that the data mixture and the order in which samples are visited during training can significantly influence model accuracy. We build and present Mixtera, a data plane for foundation model training that enables users to declaratively express which data samples should be used in which proportion and in which order during training. Mixtera is a centralized, read-only layer that is deployed on top of existing training data collections and can be declaratively queried. It operates independently of the filesystem structure and supports mixtures across arbitrary properties (e.g., language, source dataset) as well as dynamic adjustment of the mixture based on model feedback. We experimentally evaluate Mixtera and show that our implementation does not bottleneck training and scales to 256 GH200 superchips. We demonstrate how Mixtera supports recent advancements in mixing strategies by implementing the proposed Adaptive Data Optimization (ADO) algorithm in the system and evaluating its performance impact. We also explore the role of mixtures for vision-language models.

cs.LG↗

arXiv2025

Apertus: Democratizing Open and Compliant LLMs for Global Language Environments

We present Apertus, a fully open suite of large language models (LLMs) designed to address two systemic shortcomings in today's open model ecosystem: data compliance and multilingual representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting `robots.txt` exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of memorization, we adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. The Apertus models also expand multilingual coverage, training on 15T tokens from over 1800 languages, with ~40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivalling or surpassing open-weight counterparts. Beyond model weights, we release all scientific artifacts from our development cycle with a permissive license, including data preparation scripts, checkpoints, evaluation suites, and training code, enabling transparent audit and extension.

cs.CL↗

arXiv2025

On Distributed Larger-Than-Memory Subset Selection With Pairwise Submodular Functions

Modern datasets span billions of samples, making training on all available data infeasible. Selecting a high quality subset helps in reducing training costs and enhancing model quality. Submodularity, a discrete analogue of convexity, is commonly used for solving such subset selection problems. However, existing algorithms for optimizing submodular functions are sequential, and the prior distributed methods require at least one central machine to fit the target subset in DRAM. At billion datapoint scale, even the subset may not fit a single machine, and the sequential algorithms are prohibitively slow. In this paper, we relax the requirement of having a central machine for the target subset by proposing a novel distributed bounding algorithm with provable approximation guarantees. The algorithm iteratively bounds the minimum and maximum utility values to select high quality points and discard the unimportant ones. When bounding does not find the complete subset, we use a multi-round, partition-based distributed greedy algorithm to identify the remaining subset. We discuss how to implement these algorithms in a distributed data processing framework and empirically analyze different configurations. We find high quality subsets on CIFAR-100 and ImageNet with marginal or no loss in quality compared to centralized methods, and scale to a dataset with 13 billion points.

cs.LG↗

arXiv2025

Modyn: Data-Centric Machine Learning Pipeline Orchestration

In real-world machine learning (ML) pipelines, datasets are continuously growing. Models must incorporate this new training data to improve generalization and adapt to potential distribution shifts. The cost of model retraining is proportional to how frequently the model is retrained and how much data it is trained on, which makes the naive approach of retraining from scratch each time impractical. We present Modyn, a data-centric end-to-end machine learning platform. Modyn's ML pipeline abstraction enables users to declaratively describe policies for continuously training a model on a growing dataset. Modyn pipelines allow users to apply data selection policies (to reduce the number of data points) and triggering policies (to reduce the number of trainings). Modyn executes and orchestrates these continuous ML training pipelines. The system is open-source and comes with an ecosystem of benchmark datasets, models, and tooling. We formally discuss how to measure the performance of ML pipelines by introducing the concept of composite models, enabling fair comparison of pipelines with different data selection and triggering policies. We empirically analyze how various data selection and triggering policies impact model accuracy, and also show that Modyn enables high throughput training with sample-level data selection.

cs.LG↗

arXiv2022

Efficiently Computing Directed Minimum Spanning Trees

Computing a directed minimum spanning tree, called arborescence, is a fundamental algorithmic problem, although not as common as its undirected counterpart. In 1967, Edmonds discussed an elegant solution. It was refined to run in $O(\min(n^2, m\log n))$ by Tarjan which is optimal for very dense and very sparse graphs. Gabow et al.~gave a version of Edmonds' algorithm that runs in $O(n\log n + m)$, thus asymptotically beating the Tarjan variant in the regime between sparse and dense. Despite the attention the problem received theoretically, there exists, to the best of our knowledge, no empirical evaluation of either of these algorithms. In fact, the version by Gabow et al.~has never been implemented and, aside from coding competitions, all readily available Tarjan implementations run in $O(n^2)$. In this paper, we provide the first implementation of the version by Gabow et al.~as well as five variants of Tarjan's version with different underlying data structures. We evaluate these algorithms and existing solvers on a large set of real-world and random graphs.

cs.DS↗

arXiv2022

What's Wrong with Deep Learning in Tree Search for Combinatorial Optimization

Combinatorial optimization lies at the core of many real-world problems. Especially since the rise of graph neural networks (GNNs), the deep learning community has been developing solvers that derive solutions to NP-hard problems by learning the problem-specific solution structure. However, reproducing the results of these publications proves to be difficult. We make three contributions. First, we present an open-source benchmark suite for the NP-hard Maximum Independent Set problem, in both its weighted and unweighted variants. The suite offers a unified interface to various state-of-the-art traditional and machine learning-based solvers. Second, using our benchmark suite, we conduct an in-depth analysis of the popular guided tree search algorithm by Li et al. [NeurIPS 2018], testing various configurations on small and large synthetic and real-world graphs. By re-implementing their algorithm with a focus on code quality and extensibility, we show that the graph convolution network used in the tree search does not learn a meaningful representation of the solution structure, and can in fact be replaced by random values. Instead, the tree search relies on algorithmic techniques like graph kernelization to find good solutions. Thus, the results from the original publication are not reproducible. Third, we extend the analysis to compare the tree search implementations to other solvers, showing that the classical algorithmic solvers often are faster, while providing solutions of similar quality. Additionally, we analyze a recent solver based on reinforcement learning and observe that for this solver, the GNN is responsible for the competitive solution quality.

cs.LG↗

arXiv2021

Law Smells: Defining and Detecting Problematic Patterns in Legal Drafting

Building on the computer science concept of code smells, we initiate the study of law smells, i.e., patterns in legal texts that pose threats to the comprehensibility and maintainability of the law. With five intuitive law smells as running examples - namely, duplicated phrase, long element, large reference tree, ambiguous syntax, and natural language obsession -, we develop a comprehensive law smell taxonomy. This taxonomy classifies law smells by when they can be detected, which aspects of law they relate to, and how they can be discovered. We introduce text-based and graph-based methods to identify instances of law smells, confirming their utility in practice using the United States Code as a test case. Our work demonstrates how ideas from software engineering can be leveraged to assess and improve the quality of legal code, thus drawing attention to an understudied area in the intersection of law and computer science and highlighting the potential of computational legal drafting.

cs.IR↗

arXiv2020

Maps for Learning Indexable Classes

We study learning of indexed families from positive data where a learner can freely choose a hypothesis space (with uniformly decidable membership) comprising at least the languages to be learned. This abstracts a very universal learning task which can be found in many areas, for example learning of (subsets of) regular languages or learning of natural languages. We are interested in various restrictions on learning, such as consistency, conservativeness or set-drivenness, exemplifying various natural learning restrictions. Building on previous results from the literature, we provide several maps (depictions of all pairwise relations) of various groups of learning criteria, including a map for monotonicity restrictions and similar criteria and a map for restrictions on data presentation. Furthermore, we consider, for various learning criteria, whether learners can be assumed consistent.

cs.LG↗

arXiv2020

Learning Languages with Decidable Hypotheses

In language learning in the limit, the most common type of hypothesis is to give an enumerator for a language. This so-called $W$-index allows for naming arbitrary computably enumerable languages, with the drawback that even the membership problem is undecidable. In this paper we use a different system which allows for naming arbitrary decidable languages, namely programs for characteristic functions (called $C$-indices). These indices have the drawback that it is now not decidable whether a given hypothesis is even a legal $C$-index. In this first analysis of learning with $C$-indices, we give a structured account of the learning power of various restrictions employing $C$-indices, also when compared with $W$-indices. We establish a hierarchy of learning power depending on whether $C$-indices are required (a) on all outputs; (b) only on outputs relevant for the class to be learned and (c) only in the limit as final, correct hypotheses. Furthermore, all these settings are weaker than learning with $W$-indices (even when restricted to classes of computable languages). We analyze all these questions also in relation to the mode of data presentation. Finally, we also ask about the relation of semantic versus syntactic convergence and derive the map of pairwise relations for these two kinds of convergence coupled with various forms of data presentation.

cs.LO↗

arXiv2020

A Strategic Routing Framework and Algorithms for Computing Alternative Paths

Traditional navigation services find the fastest route for a single driver. Though always using the fastest route seems desirable for every individual, selfish behavior can have undesirable effects such as higher energy consumption and avoidable congestion, even leading to higher overall and individual travel times. In contrast, strategic routing aims at optimizing the traffic for all agents regarding a global optimization goal. We introduce a framework to formalize real-world strategic routing scenarios as algorithmic problems and study one of them, which we call Single Alternative Path (SAP), in detail. There, we are given an original route between a single origin--destination pair. The goal is to suggest an alternative route to all agents that optimizes the overall travel time under the assumption that the agents distribute among both routes according to a psychological model, for which we introduce the concept of Pareto-conformity. We show that the SAP problem is NP-complete, even for such models. Nonetheless, assuming Pareto-conformity, we give multiple algorithms for different variants of SAP, using multi-criteria shortest path algorithms as subroutines. Moreover, we prove that several natural models are in fact Pareto-conform. The implementation of our algorithms serves as a proof of concept, showing that SAP can be solved in reasonable time even though the algorithms have exponential running time in the worst case.

cs.DS↗