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.
Zhen Huang, Yikun Wang, Shijie Xia et al.· 0 citations
Extensive context has become the norm as Large Language Models (LLMs) are increasingly deployed in long-horizon search tasks. The concern that increasing context length degrades model capabilities, known as context rot, has become a widely recognized issue for these applications. However, in deep search scenarios, it remains unclear how models actually fail under extensive context, and to what extent existing methods can mitigate such failures. Through a systematic study of four flagship models across three benchmarks, we identify a previously overlooked phenomenon, which we term premature termination: under extensive context, models give up or provide uncertain incorrect answers long before exhausting the context window. By controlling for query difficulty, we show that the premature termination rate is positively correlated with context length. Based on the findings, we revisit methods to mitigate context rot, including context management and parallel sampling. For context management, we analyze seven methods across three categories and show that they are inherently test-time scaling strategies that reduce the premature termination rate to enable more exploration, and we further provide model-dependent principles for method selection. For parallel sampling, we develop a behavior-aware filtering strategy and observe a performance gain of 2.6% to 4.9% across three aggregation methods.