Task vectors from training on G1i mostly applied cleanly to G1j; publishing merged version here plus a touchup focusing on consolidating tool use into the chosen paradigm.
(Includes several further adapters not currently public. The 'adapter' folder is purely the touchup and does not need to be applied.)
This model should be considered an In Progress checkpoint. They are not strictly a base model, though I have included an auxiliary loss reward to preserve language modeling. However, they will not have the full capabilities of most post-trained models.
Also, RNNs have different natural capabilities than Transformers, and some tasks will be harder to learn to carry in state than attention.