Language Model Pretraining

Modern language models are trained on heterogeneous web-scale text corpora. Consequently, studying knowledge and skill acquisition is difficult, as prior exposure to related content is hard to characterize. To address this challenge, we introduce LittleCurriculum, a curated 88B-token pretraining corpus tailored to U.S. elementary school material, explicitly excluding concepts, facts, and vocabulary taught above Grade 5. Training a 5B-parameter LLM from scratch on LittleCurriculum yields LittleLearner, a model with sufficient language competence for open-ended evaluation, yet with clear knowledge and capability boundaries mapped to interpretable curriculum guidelines. We release LittleCurriculum and LittleLearner as a developmentally restricted sandbox to study how models acquire, represent, and use data under a well-defined training scope. We illustrate the sandbox's utility in a first suite of experiments on injecting new knowledge through post-training and in-context learning. These methods let LittleLearner better utilize existing knowledge, but do not raise out-of-scope capabilities. Our findings underscore the value of this controlled environment for future investigations.

Measuring training data influence consistently across language model pretraining is challenging. It is difficult to select downstream tasks or validation sets representative of a model's general capabilities, and reliance on task performance at intermediate checkpoints complicates comparisons across training. We propose a measure of training data influence that does not require selecting a downstream task or validation set as the attribution target. Specifically, we define an example's influence by how much its gradient update reduces the squared distance to the final parameters of a given pretraining run, and estimate this quantity from intermediate checkpoints without retraining. Applying the method to 18 configurations from the Pythia and PolyPythia suites, we find systematic temporal changes in influential data. Early in training, literature-related data are more strongly aligned with the trajectory toward the final parameters, whereas STEM data become more strongly aligned in later stages. This qualitative crossover is broadly consistent across model configurations. Our results provide a tractable trajectory-level view of how influential data change throughout pretraining, complementing influence analyses defined with respect to specific downstream tasks or validation sets.

We ask whether language-model pre-training can be decomposed into smaller, independently trainable jobs that can later be recomposed into a coherent larger model. We introduce Mixture of Training (MoT), a scaffolded modular pre-training procedure that partitions a target Transformer into contiguous layer blocks, trains each block inside a frozen pretrained aligner scaffold, and then recomposes the trained blocks with an optional short end-to-end adaptation pass. On a 1.3B-parameter Gemma-style model trained on C4, MoT provides a small-scale proof of mechanism: independently trained depth slices can be recomposed into a usable language model, and a quality-parity schedule reaches the same reported perplexity as the monolithic baseline. This parity setting processes more aggregate tokens and has a shorter idealized layer-equivalent critical path after aligner preparation; its effective compute advantage depends on reusing the aligner across runs. We therefore present MoT not as a general replacement for monolithic pre-training, but as a small-scale framework for studying whether scaffolded sub-runs can act as reusable training units.

Large language models (LLMs) have achieved remarkable breakthroughs across various applications. However, their architectures remain inefficient in pretraining due to two main limitations: (i) self-attention lacks an explicit inductive bias for locality, leading to redundant modeling of sequence-internal local information; (ii) mixture-of-experts (MoE) implicitly couples knowledge storage with computational pathways, hindering flexible access to sequence-external global knowledge. To overcome these limitations, we propose LoKiFormer, a novel LLM architecture that augments the standard decoder with two dedicated modules: 1) Local Fusion Attention (LFA), which incorporates a convolutional fusion to attention, explicitly capturing local patterns and allowing the attention to operate on more informative representations; 2) Knowledge Memory Module (KMM), which introduces a parametric key-value memory that explicitly stores global knowledge in addressable slots, decoupling storage from computation and enabling direct knowledge retrieval. Together, these modules enable LoKiFormer to achieve more efficient and effective integration of information at both levels. Experimental results show that LoKiFormer converges 1.33x faster in pre-training than baseline models, underscoring its superiority over existing LLM architectures.

The lack of diversity in LM content is widely attributed to the alignment process, but how and where exactly in the pipeline this collapse begins is unknown. We argue that output homogeneity is likely learned during the pretraining phase, and only revealed or magnified during the alignment process. Specifically, we find that semantic convergence is observed from the first alignment stage--the instruction-tuning phase (SFT)--suggesting that homogeneity might already exist in the pre-alignment model. To investigate this, we conduct controlled SFT experiments examining how training data influences output convergence on specific input/output pairs. We find that convergence can be revealed and amplified, but not introduced by the SFT data, supporting its role as a catalyst rather than a cause. To further test whether homogeneity originates before alignment, we measure convergence in base models. We find that instruct-like collapse can be induced through prompting alone, even without alignment. Taken together, our results suggest that semantic convergence may arise naturally from the objectives underlying LM training, making it difficult to mitigate through post-alignment interventions alone.

Pretraining corpus composition shapes LLM capabilities, but it often remains hidden even when model weights are released. Prior work has inferred corpus mixtures or traced specific token groups from released tokenizer vocabularies; in contrast, we estimate corpus ratios for arbitrary target tokens. We first show that BPE tokenizers trained on different corpora share stable token ID--ratio distributions, motivating distribution transfer from known corpora to a target tokenizer trained on hidden corpora. We then propose Quantile-Guided Density Estimation (QGDE), which approximates this distribution with multiple quantile trends and uses local density weighting to produce token-level estimates. In controlled settings and a realistic setting using the released SmolLM tokenizer, QGDE achieves mean relative errors as low as 3.00% for token-level estimation and 3.08% after aggregation into category-level mixtures. These results suggest that released tokenizer vocabularies provide a useful signal for fine-grained corpus estimation beyond coarse composition inference.

Diffusion language models (dLLMs) generate text through iterative denoising, allowing multiple tokens to be predicted in parallel. However, pretraining may mask tokens throughout a sequence, whereas prompt continuation conditions on an intact prefix. This difference remains in conversion pipelines that denoise entire sequences during the stable stage. We propose Prefix-Conditioned Diffusion (PCD), which samples a boundary, preserves the prefix, and denoises the suffix. The training recipe also applies autoregressive supervision to the prefix. We evaluate PCD in the stable stage of a warmup, stable, and decay conversion pipeline, with inference unchanged. Matched experiments across model families show improvements in reasoning and coding over native diffusion training. A matched continuation study further shows that the advantage persists after a shared decay stage. In a separate reconstruction diagnostic, the full PCD recipe's advantage over native diffusion training reverses as more evaluation prefix tokens are masked. Controlled experiments show lower suffix reconstruction loss with a clean training prefix, an intact evaluation prefix, and no autoregressive loss.

How a document's arrangement is written down, its notation, is a training variable that no dataset card records. The field has established that text-extraction choices change model behaviour, and has never once measured the notation of what those choices put into the corpus. We define clean-window survival, a deterministic count of how much of a stream still demands the boundary inference, and measure notation on three fronts. What corpora carry: a census of thirteen public corpora, where survival falls to 0.153 in a vision-converted PDF slice against 0.889 in C4; the scarce resource is not unmarked text but long unmarked text; a pre-registered supply test finds what remains institutional, not consumer. Our own pre-registered prediction failed: converters do not fabricate structure on prose, and that null forced the reliability mechanism that survives it. What readers use: across five base models spanning 0.6B to 8.2B and two pipelines, deleting a structural announcement makes the following prose measurably harder to predict, while swapping its notation moves nothing. That zero does not make notation unimportant; it relocates the variable: the operative cue is the announcement, not the sigil. What writers impose: a bounded null. Base models do not impose the marked register above the authored baseline, and handed prose with every announcement deleted they do not put one back, at a rate indistinguishable from zero against an authored reference of zero. We ship the format those measurements imply: the pure frame, paragraphs in authored order, every announcement deleted into a reversible sidecar, mixed against the marked copy over announcement presence rather than notation. Choose format operators by the capability they train, not by the fidelity they preserve, and record extractor identity and survival on data cards.

Pretraining LLMs on artificial languages ("pre-pretraining") is a technique that could reportedly increase token efficiency by 33%, i.e., save up to 33% of training tokens needed to reach a certain performance. We validate this prior result for English on a larger set of natural languages across four language families, using two different tokenizers and varying model sizes. We also relate the observed gains (or losses) in token efficiency to quantified linguistic properties of the languages, such as sentence length, morphological richness, and features of dependency syntactic trees (tree depth, number of children, number of crossing dependencies). Our empirical results indicate that the reported gains depend heavily on the experiment setup and the choice of random seed, although we can confirm the trend of stable gains with 128-Dyck pretraining of small models with the Llama tokenizer for most of the examined languages. On a general note, we argue that multiple training runs should be carried out at least for a subset of experiments to avoid the community adopting unstable approaches.

Pretrained Transformer encoders such as AraBERT, MARBERT, and CAMeLBERT have become the standard backbone for Arabic natural language understanding, but their self-attention mechanism scales quadratically with sequence length, which limits efficiency on long documents. Mamba, a selective state-space model (SSM), offers linear-time sequence modeling as a competitive alternative to attention, yet no dedicated bidirectional Mamba encoder pretrained specifically for Arabic currently exists. We introduce AraSSM, a bidirectional Mamba encoder pretrained via masked language modeling on a corpus combining Arabic Wikipedia and CulturaX text, trained end-to-end on four consumer-grade NVIDIA RTX 2080Ti GPUs (11GB) over approximately ten days. We evaluate AraSSM by fine-tuning on four established Arabic NLU benchmarks covering sentiment classification (HARD), named entity recognition (ANERcorp), extractive question answering (ARCD), and natural language inference (XNLI-ar), following the per-task evaluation protocol introduced by AraBERT, and report results as mean +/- standard deviation across three fine-tuning seeds. AraSSM matches or exceeds published base-sized Transformer baselines on sentiment classification (96.37 +/- 0.03% accuracy on HARD), is competitive on extractive QA (32.19 +/- 1.07 EM, 63.79 +/- 0.25 F1 on ARCD) and named entity recognition (81.54 +/- 0.30 entity-level F1 on ANERcorp), and trails the base-sized Transformer range on natural language inference (72.83 +/- 0.07% accuracy on XNLI-ar), despite being trained entirely from scratch on consumer hardware rather than large-scale accelerator clusters.

We test whether decoder-only language-model FFNs require SwiGLU's open positive tail. We introduce MemGLU as a closed-tail comparator derived from a memristive branch geometry. Across paired 9M and 30M pretraining runs with three seeds, MemGLU remains within about 0.1% of SwiGLU in validation NLL. Trained SwiGLU checkpoints are sensitive to positive-tail suppression, while mechanism diagnostics show that the two models use their gates differently despite similar losses. These results suggest that models adapt to the gate geometry available during pretraining. At the tested scales, SwiGLU's open positive tail is not necessary for decoder-only language-model FFNs.

Positional embeddings (PE) in Transformers encode token distance and order but are largely agnostic to syntactic structure. We introduce Syntax-informed Positional Embeddings (SiPE), which learns a lightweight syntactic prior from dependency parses during pretraining and injects it across all three dominant PE families (absolute, relative, rotary), for both encoders and decoders, leaving self-attention and the rest of the architecture untouched. We isolate where and how the prior should enter the model, and find it depends on the architecture: for autoregressive decoders that use relative PE, the prior is strongest when coupled multiplicatively with the relative-position term of the attention score, outperforming injection into the input embeddings, into self-attention, or into the positional and attention terms jointly---while for encoders it is best added directly to the input embeddings, composing with each encoder's native positional mechanism. We find that models pre-trained with SiPE improve on the SyntaxGym benchmark by up to $10.3\%$ while simultaneously reducing perplexity by $9.0\%$ over a base model with no syntactic supervision---a metric nearly every existing syntax-injection method instead degrades. Crucially, these gains extend beyond syntactic generalization: SiPE also improves real-world language understanding, raising scores on the GLUE benchmark by up to $8.2\%$ over a model trained without it. Unlike existing syntactic language models that marginalize over many parses at inference or discard syntax at runtime, SiPE conditions on a single parse, establishing a new Pareto frontier between syntactic supervision and inference cost.

We present MameLoshnLM, the first open-source 8B-parameter language model built specifically for Yiddish. Despite Yiddish's rich textual tradition, its limited digital presence and the scarcity of reliable evaluation resources have constrained progress in Yiddish language modeling. Existing multilingual corpora and benchmarks are often poor proxies for the language, containing substantial amounts of noisy, machine-translated, and misclassified text. We address these gaps by introducing Oytser, a high-quality Yiddish pretraining corpus that combines contemporary web-native sources with literary materials, and Kashes, a multi-task benchmark spanning translation, linguistic analysis, information extraction, and language understanding. Using these resources, we continue pretraining Llama 3.1 8B to obtain MameLoshnLM. Across the tasks in the benchmark, MameLoshnLM outperforms open baselines of similar scale. Our analyses show that these gains are not only quantitative: relative to general-purpose multilingual models, MameLoshnLM better captures language-defining lexical and morphological patterns, pointing to a broader failure mode of noisy web-scale multilingual data for low-resource languages. Our results provide both a foundation for Yiddish NLP and a practical template for language model development in historically rich but digitally underrepresented languages.

While standard Next-Token Prediction (NTP) lays the foundation of language model pre- training, its teacher-forced training paradigm may not be optimal for long-horizon reasoning and planning. Recent works such as Multi-Token Prediction (MTP) and Next-Latent prediction (NextLat) try to mitigate the problem through predicting multiple future tokens and self-supervised prediction in the latent space. However, those auxiliary objectives either have a limited horizon or suffer from compounding error from multi-step rollout. We introduce Hierarchical Latent Prediction (HiLP), which introduces an auxiliary higher-level abstract latent to help reduce the error accumulation effect in latent-space rollouts. Experiments show that HiLP can lead to longer-horizon coherent belief state representation and demonstrate the effectiveness of our method across coding and multi-step reasoning benchmarks, and offers more speculative decoding efficiency.

Dialectal variation remains a major challenge for multilingual language models. Perturbation-based continued pre-training (CPT) has emerged as a promising approach to improving robustness, yet existing work largely evaluates individual perturbation strategies in isolation and provides limited insight into why they work. We present a systematic study of perturbation-based CPT for multilingual dialect robustness in LLMs, comparing six training conditions across nine German, Italian, and Arabic dialect tasks. Perturbation-based CPT, especially character-noised CPT, consistently improves zero-shot dialect robustness while largely preserving standard variety performance. More importantly, we show that methods with similar downstream performance induce distinct mechanisms of robustness, exhibiting different patterns of language model adaptation, representational alignment, and prediction repair. Our results provide a more complete understanding of how synthetic surface variation improves robustness and offer practical guidance for selecting CPT strategies in multilingual and dialectal settings.

Muon has recently emerged as a promising alternative to AdamW for language model pretraining by orthogonalizing momentum matrices using Newton-Schulz iterations. Although Muon mitigates gradient anisotropy, it does not explicitly account for the curvature geometry of the loss landscape and may therefore remain sensitive to curvature anisotropy. We bridge this gap by proposing MALT (Muon Augmented by Lightweight Two-sided Preconditioning), which uses lightweight diagonal preconditioners to reduce the sensitivity of Muon to curvature anisotropy. Specifically, MALT uses two-sided diagonal preconditioners with low memory and computational overhead to approximately capture the curvature geometry of the loss landscape. It orthogonalizes the preconditioned momentum using Newton-Schulz iterations and maps the result back to define the update direction, while norm grafting controls the update magnitude. To improve the robustness of MALT to stochastic gradient noise, we further propose MALTER (MALT with Adaptive stEpsize Rescaling). Convergence guarantees are provided for MALT in the stochastic non-convex setting. Experiments on GPT-2 Small, Medium, and Large pretraining show that the proposed methods outperform Muon while maintaining nearly the same memory footprint and wall-clock time.

We identify a previously overlooked failure mode of ALiBi positional encoding: its linear bias scaling underflows floating-point precision, which zeroes out a large fraction of attention weights and renders the affected attention heads partially blind. We analyze this failure mode, characterize its impact, and examine four mitigation strategies. We further demonstrate its occurrence in state-of-the-art pretrained models based on ALiBi. Comprehensive pretraining experiments with 148M-parameter decoder models help us to disentangle its effects from out-of-context degradation. We find that ALiBi's failure mode can substantially impair token retrieval while having only a minor effect on standard decoder benchmarks. We propose four training-time mitigation strategies and evaluate them individually and in combinations, finding that log-scaled distances yield the most consistent improvements in passkey retrieval. Despite this problem, default ALiBi slopes remain a surprisingly strong baseline, particularly for needle-in-a-haystack retrieval. Based on these findings we provide concrete recommendations on how to train models with ALiBi.

Pre-pretraining language models (LMs) on symbolic data can accelerate and improve natural language acquisition. However, existing pre-pretraining tasks, such as Dyck and procedural algorithms, rely on narrow primitives that fail to capture the expressive capacity of natural language. Moreover, prior studies remain restricted to relatively small token budgets, offering limited insight into skill emergence and representational dynamics. To address these limitations, we propose logic pre-pretraining (Logic-PPT) as a principled initialization strategy, leveraging formal derivations to impart richer structural and linguistic biases. Formal derivations require abstract mechanisms that are central to natural language, simultaneously binding variables, connecting quantifiers and relational dependencies, and composing predicate-argument structures over long contexts. Scaling our evaluation to a 100B-token regime, logic pre-pretraining substantially accelerates skill acquisition in LMs, achieving 80% accuracy on linguistic tasks with 36B fewer tokens than standard initialization, and outperforming alternative pre-pretraining baselines. Mechanistically, formal derivations induce persistent structural reorganization, distinctively characterized by a lower-rank, spectrally concentrated representation space. Crucially, we show that this internal geometry enables improved model compressibility via pruning, matching the dense baseline performance even at $\approx$33% sparsity.

Large-scale pretraining corpora contain substantial duplicate content. Although document-level deduplication is widely used, removing subdocument-level redundancy remains challenging. At corpus scale, suffix-array-based methods are commonly applied independently within shards, leaving cross-shard duplicates undetected and making the resulting retention behavior sensitive to the sharding configuration. Hash-based methods enable global exact duplicate counting, but often rely on fixed copy-retention policies that cannot accommodate heterogeneous repetition patterns. We propose a scalable subdocument deduplication framework that decouples duplicate detection from copy retention. It identifies duplicate groups through natural-boundary segmentation, normalized exact hashing, and distributed aggregation, and then applies an explicit frequency- and length-aware retention policy that allocates an adaptive copy budget to each group, retaining more copies of low-frequency or short repetitions while more aggressively deleting high-frequency or long ones. Experiments on FineWeb-Edu and a code-containing web corpus show that models trained on data processed by our method achieve the best overall performance among the evaluated settings. These results underscore the importance of explicit copy-retention control.

In recent years, the rapid development of language models has transformed the field of Natural Language Processing through a wide range of applications. However, the development of language models has not progressed uniformly across all languages. In the case of the Portuguese language, there has recently been a growing effort by academia and companies to develop language models and create data resources for Portuguese. These efforts have resulted in the rise of an increasingly diverse ecosystem of language models for Portuguese. However, information on these models remains dispersed in scientific publications, technical reports, model repositories, and project documentation. This survey presents a systematic mapping study of language models developed for Portuguese, providing a comprehensive overview of the current state of the field. We map a total of 46 models, characterizing them by various aspects, including base model, architecture, computational resources, training datasets, licensing, code availability, data, and model weights. Furthermore, we analyzed the evolution and relationships among these models through a phylogenetic perspective, identified current research gaps and opportunities, and discussed future directions for the development of language models for Portuguese.

Synthetic textbook data has improved language model pre-training, but prior work largely treats the benefit as a property of generated content or local rewriting style. We study a different factor: whether related content is organized into coherent book-level documents. We contribute both a scalable synthesis pipeline and controlled evidence that this organization matters. The pipeline retrieves source material from a pre-training corpus, clusters it into topical units, plans hierarchical tables of contents, and assembles source-grounded sections into complete books (our Full setting), yielding 686K textbooks (32B tokens) across 15,000+ disciplines. Replacing natural books in a mid-training mix with this corpus improves downstream performance by +1.09 on average. Controlled comparisons then disentangle the relevant design factors. A content-matched Split condition holds generated text and tokens fixed but treats each section as an independent document; Full's +1.02 mean gain isolates document packaging. A length-matched RandomConcat control that joins sections from different books remains below Full, ruling out document length alone. A retrieval-pool-matched Rephrase condition independently rewrites individual retrieved documents under the same audience-by-style scheme, without clustering, TOC planning, or book assembly; Full's +1.17 gain demonstrates the value of structured synthesis. On Llama3-8B, Full likewise outperforms both RandomConcat and Natural Books, supporting book-level organization as a useful axis for synthetic pre-training data design.

Decoder-only language models entangle long-term memory and reasoning in a single parameter set, making it difficult to scale memory capacity independently. Memory Decoder introduces a parametric long-term memory module but only studies it at a relatively small scale. In this work, we present Memory Decoder at Scale, scaling memory models up to 6.9B parameters and pretraining them on 300B tokens. At this data scale, the combined cost of indexing and search makes a standard Faiss pipeline infeasible. We address this bottleneck with a distributed pipeline for Faiss indexing and retrieval, together with sparse, batch-wise loading of kNN distributions. Across model scales, we find that allocating more parameters to memory yields a better parameter-performance tradeoff than scaling the base model alone. On 17 benchmarks, pairing a 6.9B general memory with Pythia-410M raises its average score from 29.86 to 37.34, surpassing Pythia-12B (37.24) with 39% fewer total parameters. For Qwen3 Base models ranging from 0.6B to 14B, 1.7B domain memories improve the average score across the three domains by more than 9 points at every scale. Overall, our results demonstrate that independently scaling pretrained memory offers a more parameter efficient path to improving language model performance.

Classical compute-optimal scaling laws assume an unbounded supply of fresh pretraining data, yet pretraining is increasingly entering a regime in which compute grows faster than the availability of high-quality data. We propose Compute-Data (CD) scaling laws, a unified framework that bridges compute-optimal scaling, where data scales freely with compute, and data-optimal scaling, where the corpus is fixed while compute can grow without bound. CD scaling extends classical scaling laws by introducing a token-effectiveness function, $η$, which quantifies the value of a derived token-produced, for example, through multi-epoch repetition or paraphrasing-relative to a fresh token, ranging from a perfect substitute to having no value. We fit $η$ for two data-expansion strategies, multi-epoch repetition and paraphrasing, across model sizes from 14M to 600M parameters using the Dolma-3 corpus. We find that token effectiveness is far from constant: it depends jointly on model size, the tokens-per-parameter ratio, and the amount of derived data, and it saturates as the corpus is expanded. The functional form of $η$ implies diminishing returns when substituting compute for data as either model size or data availability increases. It also partitions training into three operational regimes---compute-bound, data-bound, and model-bound---and shows that classical compute-optimal allocation is suboptimal across most practically relevant settings.

Developers judge a model checkpoint by how it behaves. After supervised fine-tuning (SFT), two checkpoints that perform about the same across relevant benchmarks are treated as interchangeable, equally ready for the next alignment stage, typically preference optimization. We ask whether this judgment misses a pretraining imprint: a difference that no post-SFT benchmark reveals, yet that decides how each checkpoint responds to further training. To find out, we run a controlled experiment on the final window of pretraining, the last data trained on before instruction tuning. Six branches fork from one partially pretrained checkpoint and differ only in this window: 500 million tokens, 0.1% to 1% of the tokens that precede it. Each branch trains its window on a single data source: generic web text, filtered web text, normative discourse, safety text, mathematical text, or synthetic educational text. SFT and post-training are then identical. After SFT the branches behave near-identically, within about one point on instruction following, refusal, and capability, yet the same post-training carries them to very different endpoints, under both a direct preference optimization update and a reinforcement learning update with a verifiable reward. We measure this deviation through refusal of harmful requests: when post-training begins the safety text branch refuses no more than the web text branch, yet by the end it has lost far less of its refusal. The other four branches gain little or no protection, so the effect is selective to what the window contained. The protection requires the safety text to arrive last rather than earlier in pretraining, and it reproduces on a second model family. What a model is pretrained on last shapes how it reacts to alignment. Therefore, a checkpoint should not be evaluated by its post-SFT behavior alone, and what it was trained on last should be reported with it.

Pretraining data processing is critical to the downstream performance of Large Language Models (LLMs). However, many existing approaches define a fixed processing strategy at the corpus or domain level and apply it uniformly to many examples, without adapting to the needs of each example. We propose DataOrchestra, a framework that unifies different processing operations and orchestrates an example-specific pipeline for each example. Given a chunk of pretraining data, an orchestrator decides whether to drop, untouch, or clean it. For a chunk to be cleaned, it selects one or more downstream operations, ranging from programmatic editing to different forms of LLM-based rewriting. For each rewriting step, it further generates a concrete instruction, which is executed by the corresponding downstream tool model. We pretrain models from 0.5B to 7B from scratch on web data processed by DataOrchestra and observe stable average gains over individual data-processing methods across 11 benchmarks. DataOrchestra is also effective for math continued pretraining and outperforms stronger processing baselines, while reducing processing compute by skipping unnecessary downstream operations.

Due to the complexity of neural network loss landscapes, optimization theory is forced to rely on idealized models, and there is generally a tradeoff between how theoretically tractable the model is, and how accurately it describes the true optimization dynamics. In this work, we stress test the simplest possible model of optimization -- the quadratic model -- and show that it can be surprisingly predictive in an LLM setting with 150M parameters and 3B training tokens. Specifically, we show that Taylor expanding the model and the loss function at intermediate checkpoints through training can accurately predict the optimization dynamics over windows that can last up to 10% of training. Having established this agreement, we then turn to analyzing the structure of these local quadratic optimization problems through two lenses: the Hessian spectrum and local stability. Using Lanczos quadrature with extremely deep probes, we are able to estimate the Hessian spectrum deep into the tail, and we find a surprising amount of structure in both the eigenvalues and eigenvectors, which depends on the batch size, preconditioner, and training time. We also empirically test local linear stability at intermediate checkpoints and compare it to theoretical predictions to demonstrate that optimization in LLMs typically occurs at a stochastic edge of stability, whose nature is also determined by batch size. Our results indicate the quadratic model may be a theoretically tractable proxy for pretraining optimization dynamics.

We present OntoBook, a method that converts medical ontology structure into pretraining signal for encoder language models. Our approach has three stages: random walks through ontology graphs capture hierarchical and causal relations between medical codes, a large language model reformulates these walks into fluent textbook-style prose, and the resulting text is used to train ModernCamemBERT, a 149M-parameter French encoder, with two objectives on the same data: masked language modeling and relation prediction between code pairs. On three French medical coding benchmarks (FRACCO, Cantemist-FR, Distemist-FR), OntoBook achieves significant improvements over MLM-only pretraining, with +2.5 micro-F1 on FRACCO and +8.0 micro-F1 on Distemist. We find that alignment between objectives is necessary: misaligned training, where each task uses different data, causes a 30-point degradation. We release 1.3 million LLM-reformulated medical textbooks across three French ontologies (CIM-10, CCAM, ATC) and pretrained model checkpoints.

Federated pre-training offers a way to train foundation models on private or distributed data without centralizing the underlying datasets. However, evaluating federated pre-training remains challenging because differences in client participation and local data availability can make directly comparable evaluation difficult. Moreover, pre-training test perplexity is tied to the pre-training distribution, while downstream benchmarks introduce task-specific adaptation that may not faithfully reflect the test perplexity established during pre-training. In this work, we study which evaluation protocol more reliably reflects federated pre-training quality. Using a controlled set of centralized and federated-trained models of a 16M parameter transformer model trained on identical client data, we assess evaluation protocols by whether they preserve a reference ranking established on the same pre-training testset. We compare downstream fine-tuning on GLUE, including full, head-only, and reduced-data variants, with next-token prediction on GLUE text as an intrinsic evaluation signal. Our results show that downstream fine-tuning does not reliably preserve the pre-training ranking, whereas direct next-token prediction exhibits a strong correspondence with the pre-training test perplexity. These findings suggest that downstream fine-tuning alone can be misleading when comparing federated pre-trained models, and that evaluation signals closer to the original pre-training objective deserve greater attention.

OpenLanguageModel (OLM) is an open-source PyTorch library for building and pretraining small language models while keeping their machinery visible. In OLM, model code reads like the architecture: components are ordinary modules, while Block, Residual, Repeat, and Parallel describe how they are wired. The resulting model can move unchanged from a teaching notebook to a complete pretraining run or a research ablation. OLM connects this readable model layer to tokenizers, local and streaming datasets, optimization, mixed precision, callbacks, checkpoints, and hardware-aware CPU, single-GPU, and single-node multi-GPU execution. We demonstrate the full path by tracing GPT-2 from diagram to code, launching a FineWeb-Edu training script, replacing one attention component, and letting AutoTrainer configure the available machine. The package includes 27 presets across nine familiar model families and documentation that progresses from LM fundamentals to architecture research. Validation shows close agreement with independent reference implementations, 90.6% four-GPU weak-scaling efficiency for a 348M-parameter workload, compact architecture edits, and positive early usability results. OLM is MIT-licensed and available through PyPI, GitHub, and its documentation site.

Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it. As a result, two basic questions remain open: (1) how do pretraining choices (model size, data) shape the returns to RL compute, and (2) what does RL actually do to the model? These questions are difficult to study in the standard LLM setting: pretraining corpora are vast and uncontrolled, making it hard to attribute behaviors to pretraining versus RL, and systematic compute sweeps across both stages are prohibitively expensive. To address these challenges, we use chess as a controlled testbed for studying reasoning across the full pretraining-to-post-training pipeline. We follow the standard LLM training pipeline by pretraining language models from 5M to 1B parameters on human chess games, supervised fine-tuning on synthetic reasoning traces, and running RL on chess puzzles with verifiable rewards. Using this framework, we find that the post-RL performance at given RL compute level is well-predicted from the pretraining loss, and slope of the RL reward curves improves approximately linearly with the pretraining tokens. Beyond scaling, we find that RL does not simply sharpen the SFT policy: on easy puzzles it amplifies correct moves the SFT policy already preferred, while on hard puzzles it surfaces correct moves that were nearly absent under SFT. We further test whether our findings transfer beyond chess by training a 1B language model on math-domain text, where the same predictive pattern emerges: longer-pretrained checkpoints reach higher post-RL performance and improve faster under RL. In sum, we provide a quantitative account of the pretraining-to-RL interface and a controlled testbed for studying the science of reasoning across the full pretraining-to-post-training pipeline.