fix: remove local path again

update(weights): train rand init model for 5k steps on cosmopedia

fix(model): update module name for qk norm fields for consistency with phi/persimmon

fix(model): update module name for qk norm fields for consistency with phi/persimmon

update(config): use `bfloat16`

update(config): use `qk_layerorm` field for consistency with `transformers`

Upload StableLmForCausalLM