Hotdog-27B 二分类模型在基准上击败 Jev
作者开源了专做 0/1 分类的 Hotdog-27B，在自建二分类基准上超过 Jev，并公开训练代码与评测。
Introducing Hotdog-27B, a specialized bipartite classifier. Hotdog-27B beats jev on our independent binary classification benchmark. I expect long term binary classification will have its own set of specialized open models, served on http://classifier.dev. A significant percentage of global tasks are just 0/1 classification. Is there a bug here? Is this a hotdog? Is this document relevant? Is this coding trajectory valuable? The real question is whether by focusing on just binary classification, we can meaningfully improve either accuracy, latency, or throughput. All training code and benchmarks are public to encourage progress in this area: https://huggingface.co/opensporks/hotdog-27B https://github.com/mrmps/tinker-binary#benchmarks