chavinlo/alpaca-native

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):

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"

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	41.96
ARC (25-shot)	52.3
HellaSwag (10-shot)	77.09
MMLU (5-shot)	41.6
TruthfulQA (0-shot)	37.58
Winogrande (5-shot)	69.46
GSM8K (5-shot)	1.44
DROP (3-shot)	14.23