Baichuan Model baichuan3 - baichuan3

Full name of the model baichuan3 Model abbreviation baichuan3 Model Type Basic Large Model Release date 2024-01-29 Pre-training file size 0GB Whether it supports Chinese (Chinese optimization) Yes Maximum supported context length 8K Number of model parameters (100 million) 0.0 Model code open source protocol is not open source Pre-training results Open source commercial situation No open source-No open source Model GitHub link is not available Model HuggingFace Link No link exists Introduction to the DataLearnerAI model Published paper name (official blog name) Baichuan Intelligent Public Account: Baichuan Intelligent released a model of over 100 billion, and its Chinese evaluation surpasses GPT-4 Paper address (official blog address) mp.weixin.qq.com/s/YkubqYqVwkYGRmFEzQTGqQ Basic model No basic model Publishing agency Baichuan Intelligent Introduction to baichuan3 The biggest difference between the third-generation large language model released by Baichuan Intelligence and the previous two generations of Baichuan Models is that this version of the model is completely closed. The official only announced the results of internal tests and no other additional information. The officially announced comparison results of the baichuan3 model and other models in terms of text understanding: model CMMLU GAOKAO MMLU AGI-Eval BBH ARC-E ARC-C Baichuan 3 78.12 79.42 81.69 73.16 84.56 86.11 85.58 GPT-4 70.33 65.99 86.40 63.42 86.70 95.40 93.60 GPT-3.5 54.06 46.90 68.54 45.93 70.10 88.90 83.70 Baichuan3's evaluation results in mathematical logic reasoning and code generation: model GSM8K MATH Human Eval MBPP Crux-l Crux-O Baichuan 3 88.17 49.20 70.12 68.20 57.88 58.38 GPT-4 92.00 52.90 67.00 63.60 69.80 68.70 GPT-3.5 57.10 13.96 52.44 61.40 49.00 49.40 According to large model evaluation data collected by DataLearnerAI, Baichuan3 is basically second only to GPT-4 and Google Gemini Ultra in terms of text understanding and mathematical reasoning, ranking third, slightly higher than the GLM4 released by Intelligent Spectrum AI some time ago.   Score slightly lower than GLM4 in the programming field: