Alpaca, created by applying SelectIT to the Alpaca-GPT4 dataset.

Recent research has highlighted the importance of data quality in scaling large language models (LLMs). However, automated data quality control faces unique challenges in collaborative settings where sharing is not allowed directly between data silos. To tackle this issue, this paper proposes a novel data quality control technique based on the notion of data influence on the training dynamics of LLMs, that high quality data are more likely to have similar training dynamics to the anchor dataset. We then leverage the influence of the training dynamics to select high-quality data from different private domains, with centralized model updates on the server side in a collaborative training fashion by either model merging or federated learning. As for the data quality indicator, we compute the per-sample gradients with respect to the private data and the anchor dataset, and use the trace of the accumulated inner products as a measurement of data quality. In addition, we develop a quality control evaluation tailored for collaborative settings with heterogeneous medical domain data. Experiments show that training on the high-quality data selected by our method can often outperform other data selection methods for collaborative fine-tuning of LLMs, across diverse private domain datasets, in medical, multilingual and financial settings.

The pre-trained Large Language Models (LLMs) can be adapted for many downstream tasks and tailored to align with human preferences through fine-tuning. Recent studies have discovered that LLMs can achieve desirable performance with only a small amount of high-quality data, suggesting that a large portion of the data in these extensive datasets is redundant or even harmful. Identifying high-quality data from vast datasets to curate small yet effective datasets has emerged as a critical challenge. In this paper, we introduce SHED, an automated dataset refinement framework based on Shapley value for instruction fine-tuning. SHED eliminates the need for human intervention or the use of commercial LLMs. Moreover, the datasets curated through SHED exhibit transferability, indicating they can be reused across different LLMs with consistently high performance. We conduct extensive experiments to evaluate the datasets curated by SHED. The results demonstrate SHED's superiority over state-of-the-art methods across various tasks and LLMs; notably, datasets comprising only 10% of the original data selected by SHED achieve performance comparable to or surpassing that of the full datasets.

We present Nanbeige4-3B, a family of small-scale but high-performing language models. Pretrained on 23T high-quality tokens and finetuned on over 30 million diverse instructions, we extend the boundary of the scaling law for small language models. In pre-training, we design a Fine-Grained Warmup-Stable-Decay (FG-WSD) training scheduler, which progressively refines data mixtures across stages to boost model performance. In post-training, to improve the quality of the SFT data, we design a joint mechanism that integrates deliberative generation refinement and chain-of-thought reconstruction, yielding substantial gains on complex tasks. Following SFT, we employ our flagship reasoning model to distill Nanbeige4-3B through our proposed Dual Preference Distillation (DPD) method, which leads to further performance gains. Finally, a multi-stage reinforcement learning phase was applied, leveraging verifiable rewards and preference modeling to strengthen abilities on both reasoning and human alignment. Extensive evaluations show that Nanbeige4-3B not only significantly outperforms models of comparable parameter scale but also rivals much larger models across a wide range of benchmarks. The model checkpoints are available at https://huggingface.co/Nanbeige.

Due to the scarcity of high-quality data, large language models (LLMs) are often trained on mixtures of data with varying quality levels, even after sophisticated data curation. A natural approach to better leverage high-quality data is curriculum-based pretraining, where the model is trained on data sorted in ascending order of quality as determined by a quality metric. However, prior studies have reported limited improvements from such curriculum-based pretraining strategies. This work identifies a critical factor constraining these methods: the incompatibility between the ascending data quality order and the decaying learning rate (LR) schedule. We find that while curriculum-based training substantially outperforms random shuffling when using a constant LR, its advantage diminishes under standard LR decay schedules. Our experiments show this incompatibility can be mitigated by two simple strategies: (1) employing a more moderate LR decay schedule, where the final LR is only moderately smaller than the peak LR, and (2) replacing LR decay with model averaging, i.e., computing a weighted average of the final few checkpoints. By combining these strategies, we improve the average score on a suite of standard benchmarks by 1.64% over random shuffling, without additional data refinement. Validated on 1.5B-parameter models trained over 30B tokens with various data-quality metrics, our findings call for a re-evaluation of curriculum-based LLM pretraining and underscore the potential of co-designing data curricula with optimization methods.

High-quality pretraining data is the fossil fuel of large language models (LLMs), yet its reserves are running low for frontier models. In this paper, we introduce RePro, a novel web recycling method that trains a relatively small LM with reinforcement learning to generate effective and faithful rephrasings of pretraining data. Specifically, we design one quality reward and three faithfulness rewards, optimizing the LM rephraser to convert organic data into high-quality rephrasings while maintaining its core semantics and structure. In our experiment, we train a 4B rephraser to recycle 72B tokens sampled from DCLM-RefinedWeb. Pretraining results on 400M and 1.4B models demonstrate that RePro delivers 4.7%-14.0% relative accuracy gains over organic-only baseline on 22 downstream tasks. RePro also outperforms ReWire, the state-of-the-art web recycling method that prompts a 70B rephraser, as well as the organic baseline with a 4x larger data pool. Experiments with different amounts of recycled data highlight that RePro improves organic data efficiency by 2-3x. Individual and distributional analyses validate that RePro preserves more critical information and faithfully reflects the characteristics of organic data compared to prompting-based methods. Together, these results show that RePro provides an efficient and controllable path to effectively harness the fossil fuel of LLM pretraining. We open-source our code, rephraser, and recycled data at https://github.com/cxcscmu/RePro.

Experiments show that training on the high-quality data selected by our method can often outperform other data selection methods for collaborative fine-tuning of LLMs, across diverse private domain datasets, in medical, multilingual and financial settings.

Alpaca, created by applying SelectIT to the Alpaca-GPT4 dataset.

GPU kernel generation by LLMs has recently experienced rapid development, leveraging test-time scaling and reinforcement learning techniques. However, a key challenge for kernel generation is the scarcity of high-quality data, as most high-quality kernels are proprietary and not open-source. This challenge prevents us from leveraging supervised fine-tuning to align LLMs to the kernel generation task. To address this challenge, we develop a pipeline that generates and curates high-quality CUDA kernels with reasoning traces, motivated by a critical observation that concise yet informative reasoning traces result in robust generation of high-performance kernels. Using this pipeline, we construct our dataset ConCuR and introduce our model KernelCoder, which is the first model trained on a cu-rated dataset consisting of PyTorch, reasoning, and CUDA kernel pairs, to our knowledge. In the KernelBench setup, our model achieves significant improvements over the existing top-performing model, QwQ-32B, and outperforms all open-source models fine-tuned for kernel generation, as well as frontier models such as DeepSeek-V3.1-Think Finally, we show that the average reasoning length can serve as a metric to assess the difficulty of kernel generation tasks. The observations, metrics, and our data collection and curation pipeline can help obtain better data in the kernel generation task in the future. High-performance GPU kernels are critical to high performance in modern machine learning systems (Dao et al., 2022). However, developing them remains a costly and time-consuming task, demanding knowledge of domain-specific programming languages such as CUDA (Nickolls et al., 2008), Triton (Tillet et al., 2019), and ThunderKittens (Spector et al., 2025), as well as expertise in computer architectures. Therefore, tools are emerging to help developers develop GPU kernels. Previously, compilers like TVM (Chen et al., 2018) were introduced to generate program kernels.