alpaca-lora-7b-reproduced

This repo reproduced tloen/alpaca-lora-7b fit on the Stanford Alpaca dataset.

4x H100 training for about 1h15min, details in W&B link, there is a hyperparameter of val_set_size=500

4 x 4090 training for about 4h35min, details in W&B link, all key hyperparameters are the same

To optimize the running speed, I change these code

• load_in_8bits=False to use 16bit finetune
• comment model = prepare_model_for_int8_training to not turn some parameters to fp32 and turn off gradient checkpointing
• for 4090 enable gradient checkpointing, add model.gradient_checkpointing_enable() and model.enable_input_require_grads()

This version of the weights was trained with the following hyperparameters:

• Epochs: 10 (load from best epoch)
• Batch size: 128
• Cutoff length: 512
• Learning rate: 3e-4
• Lora r: 16
• Lora target modules: q_proj, k_proj, v_proj, o_proj

That is:

python finetune.py \
    --base_model='decapoda-research/llama-7b-hf' \
    --num_epochs=10 \
    --cutoff_len=512 \
    --group_by_length \
    --val_set_size=500 \
    --output_dir='./alpaca-lora-train-H100-80G-HBM3x4-mb8' \
    --lora_target_modules='[q_proj,k_proj,v_proj,o_proj]' \
    --lora_r=16 \
    --micro_batch_size=8 \
    --train_in_8bit False

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