---
license: mit
datasets:
- yahma/alpaca-cleaned
- teknium/GPT4-LLM-Cleaned
- databricks/databricks-dolly-15k
---


This repo contains a low-rank adapter for LLaMA-13b
fit on the [Stanford Alpaca](https://github.com/tatsu-lab/stanford_alpaca) dataset.

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

- Epochs: 10 (load from best epoch)
- Batch size: 128
- Cutoff length: 1024
- Learning rate: 2e-5
- Lora _r_: 16
- Lora target modules: q_proj, k_proj, v_proj, o_proj


That is trained by using RTX 3090 * 8 pcs around 10 hrs.:

```bash
WORLD_SIZE=8 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 nohup torchrun --nproc_per_node=8 --master_port=1234 finetune.py \
    --base_model 'decapoda-research/llama-13b-hf' \
    --data_path './alpaca_data_gpt4_dolly15k.json' \
    --output_dir './lora-alpaca-13B-gpt4-dolly15k' \
    --batch_size 128 \
    --micro_batch_size 4 \
    --num_epochs 10 \
    --learning_rate 2e-5 \
    --cutoff_len 1024 \
    --val_set_size 2000 \
    --lora_r 4 \
    --lora_alpha 16 \
    --lora_dropout 0.05 \
    --lora_target_modules '[q_proj,k_proj,v_proj,o_proj]' \
    --train_on_inputs \
    --group_by_length \
    &

```

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