6.7m (6,700,128) param GPT-J model.

n_positions - 128
n_embd - 64
n_layer - 4
n_head - 8
rotary_dim - 64
tokenizer - gpt-j

First, trained on 4,194,304 samples from the c4 dataset, at a length of 128 tokens each, that comes out to 536,870,912 (0.53B) tokens seen during training. A batch size of 16 with 128 gradient accumulation steps was used, making the effective batch size 2048. A cosine learning rate schedule was used starting at 1e-3.