On September 4, 2024, the Beijing ZhiYuan Artificial Intelligence Research Institute (BAAI) announced the launch of the world's first model combat evaluation service featuring text-to-video, called FlagEval Large Model Arena. This service is open to users and covers approximately 40 large models from both domestic and international sources, supporting customizable online or offline evaluations for four major tasks: language Q&A, multimodal image and text understanding, text-to-image, and text-to-video.

The Zhiyuan Research Institute recently released three new vector models that excel in vector retrieval tasks and have set new best standards across multiple evaluation benchmarks. The three models are: BGE-EN-ICL: an English vector model that enhances the model's semantic expression capabilities by introducing task-related query-document examples as few-shot examples.

Beijing Zhiyuan Artificial Intelligence Research Institute, in collaboration with China Telecom Artificial Intelligence Research Institute, has launched an upgraded version of the Tele-FLM series of large models, including the 52B instruction model FLM-2-52B-Instruct and the trillion-parameter model Tele-FLM-1T. The FLM-2-52B-Instruct focuses on enhancing Chinese conversational abilities through instruction fine-tuning, achieving 90% of GPT-4's capabilities, based on the Tele-FLM-52B base model, utilizing specific datasets and parameter optimization.

ZhiYuan Research Institute has released a code generation training dataset called TACO, aimed at providing more challenging training data and evaluation benchmarks for code generation models. TACO has advantages in terms of data scale, quality, and evaluation schemes, including a larger training and testing set, diverse problem-solving answers, and fine-grained labels. Experimental results show that current popular code generation models show significant differences compared to GPT-4 in TACO evaluations, indicating that there is still room for improvement in this field. TACO is not just a challenging...

["ZhiYuan Research Institute has released the new generation multimodal foundation model Emu2, pushing the boundaries of multimodal contextual learning capabilities.", "Emu2 surpasses Flamingo-80B and IDEFICS-80B, demonstrating excellent performance in few-shot multimodal understanding tasks.", "Emu2 achieves optimal performance in multiple few-shot understanding, visual question answering, and image generation tasks.", "Emu2-Chat realizes accurate understanding of text-image instructions, while Emu2-Gen offers flexible, controllable, high-quality images."]

ZhiYuan Research Institute, in collaboration with TuoSi and ZhongKe WenGe, has jointly established the 'Chinese Internet Corpus' (CCI). This corpus has undergone strict screening and cleaning, with a data scale of 104GB, covering the period from 2001 to 2023. ZhiYuan Research Institute will continue to expand data sources and improve data processing workflows to provide more high-quality and reliable data resources. The institute has also opened up other high-quality Chinese datasets, such as WUDAO corpus, COIG, and MTP. This initiative aims to support the big data and artificial intelligence industries.

ZhiYuan Research Institute has open-sourced the JudgeLM evaluation model, which can efficiently assess various large models and provide scores. Compared to GPT-4, JudgeLM's cost is only 1/120, with a consistency rate of over 90% for evaluation results. JudgeLM can be applied in various assessment scenarios including pure text and multimodal contexts, generating scores and justifying reasons. The consistency of JudgeLM with reference answers exceeds 90%, approaching human performance. ZhiYuan Research Institute has also released datasets for training and validation samples for in-depth research on large models.

["ZhiYuan Research Institute has recently open-sourced the Uni3D model with 1 billion parameters, designed for general 3D vision tasks.", "The model can process point cloud data and has achieved breakthroughs in mainstream 3D vision tasks.", "Uni3D employs a unified Transformer architecture and introduces a multimodal alignment training method.", "The model has achieved state-of-the-art results across various 3D vision tasks.", "ZhiYuan Research Institute states that the open-source release of Uni3D will contribute to the future of 3D computing."]

ZhiYuan Research Institute has unveiled the new open-source bilingual model Wudao・Tianying 34 Billion Aquila2-34B, which excels in reasoning, generalization, and more. The institute has also released a comprehensive open-source toolkit to promote collaborative innovation in large model research. Aquila2-34B surpasses other open-source foundational models in overall capabilities, with the ZhiYuan team developing the NLPE method to enhance the model's extension capabilities.

ZhiYuan Research Institute has released the open source AI hardware evaluation engine FlagPerf v1.0. FlagPerf evaluation metrics include functional correctness, performance, resource usage, and ecosystem compatibility, supporting various training frameworks and inference engines. Multiple testing environments rigorously evaluate the participating code to ensure fair and just results. The testing code has been open-sourced, and the testing process and data are reproducible.

The ZhiYuan Research Institute has released the world's largest Chinese-English semantic vector model training dataset, MTP, with a data scale of 300 million pairs. MTP is the largest open-source dataset of Chinese-English related text pairs, providing an important foundation for training semantic vector models. The dataset includes Chinese-English text pairs from multiple sources, covering various types such as Q&A, comments, and news. The ZhiYuan Research Institute stated that this data plays a crucial role in training large models and will promote collaborative innovation in artificial intelligence. The release of this dataset is expected to address the shortage of training datasets for Chinese models.