GLM-4 with hopefully fixed context and some Alpaca eval maxxing. This checkpoint is SFT only and I'm planning to do APO, but still need to gen ~10K responses for training and finish up my anti-rep preference dataset.
32K context requires only 2GB VRAM, so you can fit Q4_K_M + 32k context on a single 24GB GPU, best in class for 32B dense models.
