Nace AI 发布 51 页报告，详解 Drex v1.5 开源权重模型训练方法
Nace AI 公开了 Drex v1.5 的完整训练细节，115 万条标注数据加置信度校准，做选择题类任务的人可以看看这份 51 页报告。
Nace AI 发布 51 页技术报告，介绍其开源权重模型 Drex 的训练过程。Drex v1.5 在 115 万条带验证答案的示例决策上训练，任务覆盖合同审读到钓鱼邮件识别两个方向。第二阶段用更大的教师模型校准 Drex 的置信度，让输出的百分比与实际正确率对齐。用户可在 Nace Console 上传文档并提问，获得每个选项的概率分数。
Nace AI has published a 51-page report on how it trained Drex, its family of open-weight AI models that pick the most likely answer from a fixed list of options.
Drex v1.5 was trained on 1.15 million example decisions with verified answers, covering tasks from reading contracts to spotting phishing emails.
In a second stage, a larger AI model taught Drex how confident to be in each answer, so its percentages match how often it is actually right.
Drex can be tested in Nace Console, where users upload documents with questions and receive a probability score for each answer.
