Cloudflare推出了Clairic，一类新型System One提取模型，能够从非结构化输入中执行提取、计算、估计和预测操作1。相比传统大语言模型，该模型提供更高的准确度、更低的延迟和更低的成本1。Clairic现已在Workers AI上以预览版形式发布，编号为@cf/cloudflare/clairic1。
该模型支持文本和图像输入，能返回置信度分数和区间估计1。每个标量答案均附带p_null值（表示答案不在输入中的概率）和置信度评分1。在受损产品理赔案例中，一个由Clairic、决策模型和语言模型组成的级联管道相比全语言模型方案快69%，处理时间从61.7秒降至19.3秒，成本则降低49%1。
在旅客数量预测应用中，Clairic在21天的美国运输安全局旅客数预测任务上表现出色，平均误差仅为3.3%，其中17个实际值落在模型给出的80%置信区间内1。Cloudflare计划在不久后发布该模型的开源权重版本1。
Cloudflare has introduced Clairic, a System One class of extraction model designed to extract, compute, estimate, and predict numerical values from unstructured inputs with greater accuracy, lower latency, and reduced costs compared to traditional large language models 1. The model is now available as a preview on Workers AI under the identifier @cf/cloudflare/clairic, with plans to release open-source weights in the near future 1.
Clairic supports both text and image inputs and outputs four categories of values: extracted, computed, estimated, and predicted 1. Each scalar answer includes a confidence score and p_null, a probability indicating whether the requested value is absent from the input 1. In a cascading pipeline combining Clairic with decision models and language models, the system demonstrated significant performance gains over a full LLM approach: processing damaged product insurance claims 69% faster, completing the task in 19.3 seconds compared to 61.7 seconds, while reducing costs by 49% 1. For a 21-day prediction task on TSA passenger numbers, the model achieved an average error of 3.3%, with 17 of the actual values falling within its 80% confidence interval 1.
评论
还没有评论，欢迎留下第一条。