Skywork series models are pre-trained on 3.2TB of high-quality multilingual (mainly Chinese and English) and code data. We have open-sourced the model, training data, evaluation data, evaluation methods, etc. 天工系列模型在3.2TB高质量多语言和代码数据上进行预训练。我们开源了模型参数,训练数据,评估数据,评估方法。
L{test} - L{train},想知道GPT-4是不是也见过了训练集,所以严格来说GPT-4也不能完全说是zero-shot ?