- Create a new conda environment
conda create --name llama3 python=3.8
. - Follow Quick Start instructions from the llama-3 repo.
We will be running generate-qa.sh
, which uses torchrun
.
- Replace
--ckpt_dir
and--tokenizer_path
with your llama-3 checkpoint and tokenizer paths. - If you are not using shard
output_28118.pkl
, then replace this filename ingenerate_qa.py
. - Select image keys in
keys.py
. The image keys must within the shard you previously selected. This is how we select images to generate QA pairs from. - Run
bash generate-qa.sh
. This will generate anoutput
folder with subdirectories named after the prompt version. Output files are namedoutput_{key}_{prompt_version}
.