updated training instructions

training_files/full-training-instructions.txt

@@ -33,17 +33,19 @@ cd ..
 
 torchrun --nproc_per_node=8 --master_port=3045 ./stanford_alpaca/train.py --model_name_or_path ./llama-7b-hf --data_path ./alpaca-7b-nativeEnhanced/training_files/alpaca-megaset-fixed.json --fp16 True --output_dir ./output_7b --num_train_epochs 3 --per_device_train_batch_size 2 --per_device_eval_batch_size 2 --gradient_accumulation_steps 16 --evaluation_strategy "no" --save_strategy "steps" --save_steps 200 --learning_rate 2e-5 --weight_decay 0. --warmup_ratio 0.03 --lr_scheduler_type "cosine" --logging_steps 1 --fsdp "full_shard auto_wrap" --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' --tf32 True
 
-# now, make sure with nano that script1.py has proper paths to everything
+# now, make sure with nano that convert-hf-to-pth-16b.py has proper paths to everything
 
 pip install -q datasets loralib sentencepiece
 pip install bitsandbytes
 
-python script1.py
+python convert-hf-to-pth-16b.py
 
 git clone https://github.com/antimatter15/alpaca.cpp
 
 cd alpaca.cpp
+
 mkdir models
+
 cd ..
 
 mv consolidated.01.pth ./alpaca.cpp/models/consolidated.00.pth