Training Long-Context Large Language Models (LLMs) is challenging, as hybrid training with long-context and short-context data often leads to workload imbalances. Existing works mainly use data packing to alleviate this issue, but fail to consider imbalanced attention computation and wasted communication overhead. This paper proposes Hierarchical Balance Packing (HBP), which designs a novel batch-construction method and training recipe to address those inefficiencies. In particular, the HBP constructs multi-level data packing groups, each optimized with a distinct packing length. It assigns training samples to their optimal groups and configures each group with the most effective settings, including sequential parallelism degree and gradient checkpointing configuration. To effectively utilize multi-level groups of data, we design a dynamic training pipeline specifically tailored to HBP, including curriculum learning, adaptive sequential parallelism, and stable loss. Our extensive experiments demonstrate that our method significantly reduces training time over multiple datasets and open-source models while maintaining strong performance. For the largest DeepSeek-V2 (236B) MoE model, our method speeds up the training by 2.4$\times$ with competitive performance. Codes will be released at https://github.com/ModelTC/HBP.

With the development of deep learning tools, the online handwritten Chinese character recognition (HCCR) performance has been greatly improved by using deep neural networks (DNNs) especially for long short-term memory (LSTM). However, DNNs suffer from large consumption of computation and storage resources, which may cause problems for service providers, such as server pressure, longer service latency and higher energy consumption. To solve these problems, we propose a framework that combines singular value decomposition (SVD) and adaptive drop weight (ADW) to accelerate and compress LSTM based models. We first build an LSTM based model that achieves an accuracy of 97.83% on the ICDAR2013 online HCCR competition dataset. After restructuring the model with SVD and ADW, it can reduce the FLOPs (floating point operations per second) of the forward process by approximately 10 times and compress the model with 1/30 of the original size with only a 0.5% decrease in accuracy. Finally, integrated with our efficient forward implementation, the recognition of an online character requires only 2.7 ms in average on a CPU with a single thread, while requiring only 0.45 MB for model storage.

The development of convolutional neural networks (CNN) has led to revolutionary progress in the resolution of the offline handwritten Chinese character recognition (HCCR) problem. As the recognition rate on a standard offline HCCR testbed is outstanding, a few samples that remain misclassified have kindled our interest. In this paper, with the help of human recognition results, we present a comprehensive analysis of the samples misclassified by a state-of-the-art CNN model. We performed the analysis based on the top-1-votes, which are obtained from the statistical analysis of human recognition results, and derived the following conclusions: (1) the majority of samples with high top-1-votes were mis-labeled. Besides, by comparing the results of human recognition with that of CNN, some limitations of CNN that provide scope for further improvement are presented; (2) in the samples with medium top- 1-votes, it is shown that the samples with different confidence level have different characteristics. Specifically, some samples could be regarded as multi-label samples; (3) the samples with low top-1- votes are either wrongly written or written extensively in cursive style, which are difficult to match their given ground-truths; (4)the relationship between writing styles and misclassifications are also introduced in the paper. We believe this work should provide some insights and brings new clues on designing new classification methods to deal with these challenging samples.