Update: Small model + Engram

A developer reported surprising coherence from a small 9b model after only 15 million tokens, attributing the success to borrowed embeddings and an LM_head down-projected from 5k to 2048 d_model. This method preserved approximately 65% of the data from a larger model's embedding, which significantly improved results compared to using pure 1 hot cross entropy that would have produced gibberish.