Large Language Models (LLMs) like GPT and Llama have sparked a revolution in the field of artificial intelligence, but efficiently training these massive models while aligning them with human values remains a challenge. Reinforcement Learning with Human Feedback (RLHF) has become an important training method for LLMs in recent years, but traditional RLHF frameworks have limitations in flexibility, efficiency, and scalability. To address these issues, the ByteDance Doubao Large Model Team has open-sourced an RLHF framework called HybridFlow.

Recently, ByteDance officially addressed rumors regarding an 'intern disrupting large model training'. The company confirmed that an intern did maliciously interfere with the model training tasks of the commercialization technology team's research project. However, this action did not affect the company's official commercial projects and online business, nor did it involve other large model businesses of ByteDance. Moreover, ByteDance pointed out that claims circulating online about 'involving over 8,000 nodes and causing losses of over ten million dollars' are seriously exaggerated. According to public reports, the intern was part of ByteDance's commercialization technology team.

At a recent video cloud technology conference, Volcano Engine unveiled a significant innovation: a video preprocessing solution for large model training. This technology has been successfully applied to the Doubao video generation model, marking a significant advancement in AI video generation technology. Volcano Engine's President, Tan Dai, emphasized that AIGC and multimodal technology are profoundly changing user experience. Leveraging experience from Douyin, Volcano Engine is actively exploring the integration of AI large models with video technology to provide comprehensive solutions for enterprises. Wang Yue, Head of Video Architecture at Douyin Group, pointed out that large model training faces numerous challenges.

The release of PaddlePaddle Framework 3.0 features a core upgrade to unified automatic parallelism technology, aimed at simplifying the distributed training process for large models and improving development efficiency. The new version supports mixed parallelism techniques from four-dimensional to five-dimensional, employing various parallel methods such as data parallelism, tensor model parallelism, pipeline parallelism, and grouped parameter slicing parallelism, effectively enhancing the training efficiency of large models. The automatic parallelism technology uses tensor slicing syntax to automatically infer distributed slicing states and add communication operators, reducing the difficulty of development. The principles of automatic parallelism include distributed tensor representation and slicing inference.

Tencent Cloud recently launched an upgraded version of the Xing Mai Network 2.0, aiming to enhance the efficiency of large-scale model training. In the previous version, the synchronization communication time for the computation results of large models accounted for over 50% of the total, resulting in low efficiency. The new version of Xing Mai Network 2.0 has been upgraded in several aspects:1. Supports networking of up to 100,000 cards in a single cluster, doubling the scale, with a 60% improv

The American social media platform Reddit has reached an agreement with Google, valued at approximately $60 million annually. Springer Publishing Group collaborates with OpenAI, becoming the first publishing entity to deeply integrate journalism with artificial intelligence technology. OpenAI has signed a cooperation agreement with Axel Springer, indicating that future large model training may involve payments to data providers. Publishing industry companies possess a wealth of electronic graphic resources and are expected to become important datasets for large model training.

ZTE has launched the latest AI server that supports large model training aimed at the training and inference needs of small to medium-sized models. ZTE has introduced the G5 series servers and plans to release the latest AI server supporting large model training this year.

Shanghai AI Lab has released the large model training toolbox XTuner, which supports various hardware adaptations, significantly lowering the training cost barrier. XTuner is compatible with multiple open-source large models and can execute tasks such as incremental pre-training and instruction fine-tuning. XTuner balances usability and configurability with a built-in standardized process for one-click training.

The Megatron-LLaMA framework, jointly developed by Taotian Group and Aicheng Technology, aims to enhance the training performance of large language models while reducing training costs and maintaining compatibility with the LLaMA community, achieving a 176% speedup in training on 32 GPUs.

Data, as a new factor of production, is forming a data industry chain. Large model training requires massive amounts of high-quality data, which is seen as one of the core competitive advantages for future large model companies. The accounting and trading of data face many challenges, such as rights confirmation and security management.