Muon is Scalable for LLM Training

Recently, the Muon optimizer based on matrix orthogonalization has demonstrated strong results in training small-scale language models, but the scalability to larger models has not been proven. We identify two crucial techniques for scaling up Muon: (1) adding weight decay and (2) carefully adjusting the per-parameter update scale. These techniques allow Muon to work out-of-the-box on large-scale training without the need of hyper-parameter tuning. Scaling law experiments indicate that Muon achieves $\sim\!2\times$ computational efficiency compared to AdamW with compute optimal training. Based on these improvements, we introduce Moonlight, a 3B/16B-parameter Mixture-of-Expert (MoE) model trained with 5.7T tokens using Muon. Our model improves the current Pareto frontier, achieving better performance with much fewer training FLOPs compared to prior models. We open-source our distributed Muon implementation that is memory optimal and communication efficient. We also release the pretrained, instruction-tuned, and intermediate checkpoints to support future research.

MoBA: Mixture of Block Attention for Long-Context LLMs

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

A Self-enhancement Approach for Domain-specific Chatbot Training via Knowledge Mining and Digest

Large Language Models (LLMs), despite their great power in language generation, often encounter challenges when dealing with intricate and knowledge-demanding queries in specific domains. This paper introduces a novel approach to enhance LLMs by effectively extracting the relevant knowledge from domain-specific textual sources, and the adaptive training of a chatbot with domain-specific inquiries. Our two-step approach starts from training a knowledge miner, namely LLMiner, which autonomously extracts Question-Answer pairs from relevant documents through a chain-of-thought reasoning process. Subsequently, we blend the mined QA pairs with a conversational dataset to fine-tune the LLM as a chatbot, thereby enriching its domain-specific expertise and conversational capabilities. We also developed a new evaluation benchmark which comprises four domain-specific text corpora and associated human-crafted QA pairs for testing. Our model shows remarkable performance improvement over generally aligned LLM and surpasses domain-adapted models directly fine-tuned on domain corpus. In particular, LLMiner achieves this with minimal human intervention, requiring only 600 seed instances, thereby providing a pathway towards self-improvement of LLMs through model-synthesized training data.

Unsupervised Parallel Corpus Mining on Web Data

With a large amount of parallel data, neural machine translation systems are able to deliver human-level performance for sentence-level translation. However, it is costly to label a large amount of parallel data by humans. In contrast, there is a large-scale of parallel corpus created by humans on the Internet. The major difficulty to utilize them is how to filter them out from the noise website environments. Current parallel data mining methods all require labeled parallel data as the training source. In this paper, we present a pipeline to mine the parallel corpus from the Internet in an unsupervised manner. On the widely used WMT'14 English-French and WMT'16 English-German benchmarks, the machine translator trained with the data extracted by our pipeline achieves very close performance to the supervised results. On the WMT'16 English-Romanian and Romanian-English benchmarks, our system produces new state-of-the-art results, 39.81 and 38.95 BLEU scores, even compared with supervised approaches.

Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing

With the success of language pretraining, it is highly desirable to develop more efficient architectures of good scalability that can exploit the abundant unlabeled data at a lower cost. To improve the efficiency, we examine the much-overlooked redundancy in maintaining a full-length token-level presentation, especially for tasks that only require a single-vector presentation of the sequence. With this intuition, we propose Funnel-Transformer which gradually compresses the sequence of hidden states to a shorter one and hence reduces the computation cost. More importantly, by re-investing the saved FLOPs from length reduction in constructing a deeper or wider model, we further improve the model capacity. In addition, to perform token-level predictions as required by common pretraining objectives, Funnel-Transformer is able to recover a deep representation for each token from the reduced hidden sequence via a decoder. Empirically, with comparable or fewer FLOPs, Funnel-Transformer outperforms the standard Transformer on a wide variety of sequence-level prediction tasks, including text classification, language understanding, and reading comprehension. The code and pretrained checkpoints are available at https://github.com/laiguokun/Funnel-Transformer.

Explainable Unsupervised Change-point Detection via Graph Neural Networks

Change-point detection (CPD) aims at detecting the abrupt property changes lying behind time series data. The property changes in a multivariate time series often result from highly entangled reasons, ranging from independent changes of variables to correlation changes between variables. Learning to uncover the reasons behind the changes in an unsupervised setting is a new and challenging task. Previous CPD methods usually detect change-points by a divergence estimation of statistical features, without delving into the reasons behind the detected changes. In this paper, we propose a correlation-aware dynamics model which separately predicts the correlation change and independent change by incorporating graph neural networks into the encoder-decoder framework. Through experiments on synthetic and real-world datasets, we demonstrate the enhanced performance of our model on the CPD tasks as well as its ability to interpret the nature and degree of the predicted changes.

Bridging the domain gap in cross-lingual document classification

The scarcity of labeled training data often prohibits the internationalization of NLP models to multiple languages. Recent developments in cross-lingual understanding (XLU) has made progress in this area, trying to bridge the language barrier using language universal representations. However, even if the language problem was resolved, models trained in one language would not transfer to another language perfectly due to the natural domain drift across languages and cultures. We consider the setting of semi-supervised cross-lingual understanding, where labeled data is available in a source language (English), but only unlabeled data is available in the target language. We combine state-of-the-art cross-lingual methods with recently proposed methods for weakly supervised learning such as unsupervised pre-training and unsupervised data augmentation to simultaneously close both the language gap and the domain gap in XLU. We show that addressing the domain gap is crucial. We improve over strong baselines and achieve a new state-of-the-art for cross-lingual document classification.

Re-examination of the Role of Latent Variables in Sequence Modeling

With latent variables, stochastic recurrent models have achieved state-of-the-art performance in modeling sound-wave sequence. However, opposite results are also observed in other domains, where standard recurrent networks often outperform stochastic models. To better understand this discrepancy, we re-examine the roles of latent variables in stochastic recurrent models for speech density estimation. Our analysis reveals that under the restriction of fully factorized output distribution in previous evaluations, the stochastic models were implicitly leveraging intra-step correlation but the standard recurrent baselines were prohibited to do so, resulting in an unfair comparison. To correct the unfairness, we remove such restriction in our re-examination, where all the models can explicitly leverage intra-step correlation with an auto-regressive structure. Over a diverse set of sequential data, including human speech, MIDI music, handwriting trajectory and frame-permuted speech, our results show that stochastic recurrent models fail to exhibit any practical advantage despite the claimed theoretical superiority. In contrast, standard recurrent models equipped with an auto-regressive output distribution consistently perform better, significantly advancing the state-of-the-art results on three speech datasets.

Large-scale Cloze Test Dataset Created by Teachers

Cloze tests are widely adopted in language exams to evaluate students' language proficiency. In this paper, we propose the first large-scale human-created cloze test dataset CLOTH, containing questions used in middle-school and high-school language exams. With missing blanks carefully created by teachers and candidate choices purposely designed to be nuanced, CLOTH requires a deeper language understanding and a wider attention span than previously automatically-generated cloze datasets. We test the performance of dedicatedly designed baseline models including a language model trained on the One Billion Word Corpus and show humans outperform them by a significant margin. We investigate the source of the performance gap, trace model deficiencies to some distinct properties of CLOTH, and identify the limited ability of comprehending the long-term context to be the key bottleneck.

Stochastic WaveNet: A Generative Latent Variable Model for Sequential Data

How to model distribution of sequential data, including but not limited to speech and human motions, is an important ongoing research problem. It has been demonstrated that model capacity can be significantly enhanced by introducing stochastic latent variables in the hidden states of recurrent neural networks. Simultaneously, WaveNet, equipped with dilated convolutions, achieves astonishing empirical performance in natural speech generation task. In this paper, we combine the ideas from both stochastic latent variables and dilated convolutions, and propose a new architecture to model sequential data, termed as Stochastic WaveNet, where stochastic latent variables are injected into the WaveNet structure. We argue that Stochastic WaveNet enjoys powerful distribution modeling capacity and the advantage of parallel training from dilated convolutions. In order to efficiently infer the posterior distribution of the latent variables, a novel inference network structure is designed based on the characteristics of WaveNet architecture. State-of-the-art performances on benchmark datasets are obtained by Stochastic WaveNet on natural speech modeling and high quality human handwriting samples can be generated as well.