A controlled experiment on the final window of pretraining, the last data trained on before instruction tuning, finds that what a model is pretrained on last shapes how it reacts to alignment, and what it was trained on last should be reported with it.
Abstract
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.
It is suggested that many reported downstream gains reflect instance and seed dependent elicitation match, rather than universally improving the attainable performance ceiling, in self-supervised fine-tuning.
Wangjin Zhou, Yizhou Zhang, Yichi Wang et al.· 0 citations
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.
Yuto Nishida, Hirokazu Kiyomaru, Yusuke Oda et al.· 0 citations
It is widely assumed that continued pretraining (CPT) on a narrow, out-of-distribution corpus such as raw biological sequence must trade away a general-purpose model’s broad competence — the “alignment tax” or catastrophic-forgetting intuition. We test this directly, without any new training, by re-analyzing three checkpoints from a single lineage of a 26B-parameter Mixture-of-Experts model (Gemma-4-26B-A4B): the instruction-tuned base, the same model after biological CPT (8.7B tokens of DNA, protein, and biomedical text), and after subsequent supervised fine-tuning (SFT). Across three independent capability axes — general knowledge/reasoning (MMLU, ARC, HellaSwag), code generation (MBPP), and biomedical knowledge (BixBench) — we find that biological CPT does not degrade the model; it lifts it: MMLU +13 points, MBPP pass@1 nearly doubles (0.33 →0.63), and BixBench discrimination rises sharply (MCC 0.23 → 0.92). The single measured regression is truthfulness (TruthfulQA 8.8 points), a small and interpretable domain drift. A clean vocabulary-expansion ablation (< 0.4 pt on every general metric) confirms the gains are attributable to CPT, not tokenizer changes. Crucially, subsequent SFT narrows the model back: all three axes fall to near-base levels, revealing a consistent division of labor — CPT re-organizes and lifts the shared capability substrate; SFT cashes it out onto target tasks. We argue this reframes biological sequence not as a competitor for a foundation model’s capacity but as a form of structured scientific data that reshapes its capability profile, and that CPT and SFT should be budgeted as complementary rather than substitutable stages. All checkpoints, evaluation code, and per-example outputs are public. Highlights A training-free re-analysis of one 26B MoE lineage isolates the effect of biological continued pretraining (CPT) from tokenizer changes and from fine-tuning. Biological CPT does not cause catastrophic forgetting; it raises general knowledge (MMLU +13 pts) and code generation (MBPP pass@1 0.33 → 0.63). CPT also makes chain-of-thought reasoning 41% shorter and near-backtrack-free while pre-serving accuracy — an effect invisible to accuracy metrics. A consistent CPT-lifts / SFT-narrows division of labor recurs across four axes, reframing biological sequence as structured scientific data that reshapes a model’s capability profile. The Bigger Picture Adapting a general-purpose AI model to a specialized domain — here, the language of DNA and proteins — is usually assumed to come at a cost: teach it biology and it forgets how to reason about everything else. This “no free lunch” intuition shapes how practitioners budget compute and whether they attempt domain adaptation at all. We test the assumption directly, and without running any new training, by comparing three snapshots of the same model taken before and after biological training. The result overturns the intuition: feeding the model raw biological sequence made it better at general knowledge, at writing code, and even changed how it reasons — producing shorter, more decisive chains of thought without losing accuracy. The gains appear during the sequence-pretraining stage and are partly given back during task-specific fine-tuning, revealing that the two stages play complementary rather than interchangeable roles. This suggests a broader principle for data-centric AI: structured scientific data — biological sequence today, and by extension code, mathematics, and chemistry — is not merely knowledge to be absorbed but a lever that reshapes what a foundation model can do.
A central hope behind process supervision is that models can expose intermediate variables that matter for their later behavior. For this to help with alignment, a scratchpad must be tied to the computation: when the model writes a state, later steps should compute from that state. To test this requirement, we use a controlled state-tracking task with a known update rule, comparing models trained to report only the final state with models trained to write intermediate states before giving the final answer. At evaluation, we edit the internal representation of one written state while leaving the visible scratchpad text fixed. Because the transition rule is known, the edit has a single correct downstream consequence. In Qwen2.5-Coder-7B, the state-writing model predicts the next phase bit implied by the edited state on 80% and 91% of held-out examples across the two task variants, while pretrained and final-answer-only controls remain near baseline. Additional controls rule out generic next-token steering and copying another continuation: the prediction depends on both the edited state and the current move. The same causal-use pattern replicates across model families. Together, these results suggest a sharper goal for scratchpad oversight: not just to make intermediate reasoning legible, but to train written states that the model uses as part of its computation.
Benjamin Shih, John Winnicki, Eric Darve· arXiv.org· 0 citations
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.
Mohammed Sabry, Sean Augenstein, Keith Rush et al.· 0 citations
A single training example's contribution to a finished model is normally estimated rather than measured, because measuring it takes two expensive full pre-training runs that differ in one row of one batch. We ran that counterfactual 24 times at a small scale. We trained 32 GPT-2 models at 124M parameters from scratch on OpenWebText, over four conditions and eight seeds. At step 200 of 9,536, at peak learning rate, we replaced one row of a 256-row batch with a fixed context injection carrying a 194-token passage. The three injected conditions are: 1. fluent prose with a corpus-attested subject, 2. fluent prose with a fabricated subject matched to it within 0.14% on full-batch gradient delta, and 3. random keyboard characters. The fourth condition is an uninjected twin. The passage is learned from one exposure and then decays. Fifty steps after injection, the arm that saw a passage predicts it better than the arm that did not by 0.039 and 0.044 nats of cross-entropy on the passage, at eight of eight seeds with p<$10^{-4}$. At the final step we do not detect that difference for either passage, at p = 0.25 and p = 0.71, against minimum detectable effects of 0.025 and 0.079 nats, nor between the two passages, at p=0.54. Every geometric measure we report is taken after that decay. Our pre-registered contrast on interpolation loss barrier is +0.0068 with p = 0.509, against a minimum detectable effect of 0.032 barrier units. Held-out cross-entropy is $-0.00044$ with p = 0.310. Per-layer centered kernel alignment does not detectably separate any condition at any layer. Weight displacement reaches 44.1% of the seed-to-seed Euclidean distance and is 92% settled by the midpoint of training, while the barrier reaches 3.0% of the seed-to-seed barrier. Those two figures sit roughly 15 times apart, and that is a lower bound. The injection relocates the model within its basin without moving it out.