alpaca-lora-13b

This repo contains a low-rank adapter for LLaMA-13b fit on the Stanford Alpaca dataset.

This version of the weights was trained on a dual RTX3090 system, powered by solar energy. Total training time was about 24 hours.
The following hyperparameters were used:

Epochs: 10
Batch size: 128
Cutoff length: 256
Learning rate: 3e-4
Lora r: 16
Lora alpha: 16
Lora target modules: q_proj, k_proj, v_proj, o_proj
That is:

OMP_NUM_THREADS=4 WORLD_SIZE=2 CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node=2 --master_port=1234 finetune.py \
--base_model='decapoda-research/llama-13b-hf' \
--data_path="yahma/alpaca-cleaned' \
--num_epochs=10 \
--output_dir='./lora-alpaca-13b-256-qkvo' \
--lora_target_modules='[q_proj,k_proj,v_proj,o_proj]' \
--lora_r=16 \
--val_set_size=0 \
--micro_batch_size=32

LR warmup was tuned to fit the first epoch.

Instructions for running it can be found at https://github.com/tloen/alpaca-lora.