File size: 1,475 Bytes
d689ebe 43108a7 d689ebe 30b0f97 d689ebe 43108a7 88071b4 cc7773c ef47142 88071b4 8446ce8 d689ebe 8446ce8 3fc930f d689ebe 43108a7 d689ebe 43108a7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 |
# Stanford Alpaca
This is a replica of Alpaca by Stanford' tatsu
Trained using the original instructions with a minor modification in FSDP mode
# Other versions:
13B: https://huggingface.co/chavinlo/alpaca-13b
13B -> GPT4 : https://huggingface.co/chavinlo/gpt4-x-alpaca
## Compute Used
Trained on 4xA100s for 6H
Donated by redmond.ai
NO LORA HAS BEEN USED, this is a natively-finetuned model, hence "alpaca-native"
If you are interested on more llama-based models, you can check out my profile or search for other models at https://huggingface.co/models?other=llama
This (MIGHT) be a quantized version of this model, but be careful: https://boards.4channel.org/g/thread/92173062#p92182396
CONFIGURATION (default except fsdp):
```shell
torchrun --nproc_per_node=4 --master_port=3045 train.py \
--model_name_or_path /workspace/llama-7b-hf \
--data_path ./alpaca_data.json \
--bf16 True \
--output_dir /workspace/output \
--num_train_epochs 3 \
--per_device_train_batch_size 4 \
--per_device_eval_batch_size 4 \
--gradient_accumulation_steps 8 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 200 \
--save_total_limit 1 \
--learning_rate 2e-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--fsdp "shard_grad_op auto_wrap" \
--fsdp_transformer_layer_cls_to_wrap 'LLaMADecoderLayer' \
--tf32 True --report_to="wandb"
``` |