Ezekiel999/AksaraLLM-20B-Instruct (smoke-test checkpoint)

This is NOT the production 20B model. It is a randomly-initialized tiny preset (2 layers, 64-dim, vocab 256) pushed from a Devin scaffolding session to validate the aksaraLLMModel.save_pretrained → HF → aksaraLLMModel.from_pretrained round-trip.

The real 20B model (42 layers, 6144-dim, vocab 131 072) must be trained from random initialisation on a TPU v5p pod using aksara-train/scripts/train_20b_pretrain.py.