license: other
There is no official 22b model, this is just a weird experiment, and any potential benefits of hacking on the architecture have not been validated in any formal manner
https://huggingface.co/chargoddard/llama2-22b-blocktriangular trained one one epoch of 52k rows of Stanford Alpaca. About 11 hours on a 3090.
I had trouble with training using the other 22b method with BLOCK_DIAGONAL=True, but with this method, this is the first time I've been able to target all modules without breaking the output.
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "gate_proj", "down_proj"]
Trained at 5e-5 with r=32. For more info see https://wandb.ai/nkpz/huggingface/runs/3oy5nbtv/workspace?workspace=user-nkpz
It's been responding coherently enough that I would need to run some objective benchmarks to determine if this is better/worse than stock llama 13b