AC2 adds custom decision-model training

Applied Compute says AC2 now supports training custom decision models. An accompanying post describes adding a trainable decision head to an open-weight language model. In a test starting from Perplexity’s pplx-decider-v1-27b, the author reports toxic-comment detection F1 rising from 0.482 to 0.677 using 50,000 sampled training examples and 20,000 sampled evaluation examples, and says training took less than an hour on four B300 GPUs.