gdn-120m-fixed-128

120M gated DeltaNet (head_dim 128) language model, 20k steps of Pile pretraining, published so its recall evaluations can be reproduced.

• trained from experiment config: mlp_mixer/gdn_paper_120m_hdim128
• checkpoint file: final.ckpt (weights only; optimizer state stripped)
• global_step 20000

Part of a state-width sweep evaluated on FDA/SWDE/SQuAD, RULER NIAH, and a distance-controlled retention sweep.