[P1] Question about the dataset

[csaiedu]

In the tutorials,
How do you actually go about generating thise model weights?

lsreft = torch.load(f"../results/prod_{model_name}_{layer}_concept16k_lsreft/train/LsReFT_weight.pt")
lsreft_metadata = load_jsonl(f"../results/prod_{model_name}_{layer}_concept16k_lsreft/train/metadata.jsonl")

thanks

[frankaging]

Hi, once you download the dataset from our HuggingFace repo here, you can train them using commands like:

torchrun --nproc_per_node=4 --master_port=30000 axbench/scripts/train.py \
  --config axbench/sweep/wuzhengx/2b/l20/16k_lsreft.yaml \
  --dump_dir axbench/results/prod_2b_l20_concept16k_lsreft \
  --overwrite_data_dir axbench/concept16k/prod_2b_l20_v1/generate \
  --run_name official

the yaml files are all released. see: https://github.com/stanfordnlp/axbench/blob/main/axbench/sweep/wuzhengx/2b/l20/16k_lsreft.yaml.