mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
Hamish IvisonandGitHub
ba10af976d
- Adds dpo training script and example bash scripts. - I've confirmed that the full-finetuning script replicates tulu-v2-dpo-7b. PEFT works with it too, although I haven't done any hyperparameter tuning or trained a model using it. - Inline the beaker docker container into the dockerfile to make it portable out of Ai2. - Update readme for Tulu 2 and associated paper
44 lines
1.4 KiB
Bash
Executable File
44 lines
1.4 KiB
Bash
Executable File
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
|
|
|
|
NUM_GPUS=8
|
|
BATCH_SIZE_PER_GPU=1
|
|
TOTAL_BATCH_SIZE=128
|
|
GRADIENT_ACC_STEPS=$(($TOTAL_BATCH_SIZE/$NUM_GPUS/$BATCH_SIZE_PER_GPU))
|
|
echo "Training model using $NUM_GPUS GPUs, $BATCH_SIZE_PER_GPU batch size per GPU, $GRADIENT_ACC_STEPS gradient accumulation steps"
|
|
|
|
# Lora training
|
|
accelerate launch \
|
|
--num_machines 1 \
|
|
--num_processes $NUM_GPUS \
|
|
open_instruct/dpo_tune.py \
|
|
--model_name_or_path allenai/tulu-2-7b \
|
|
--use_qlora \
|
|
--use_lora \
|
|
--use_flash_attn \
|
|
--lora_rank 64 \
|
|
--lora_alpha 16 \
|
|
--lora_dropout 0.1 \
|
|
--tokenizer_name allenai/tulu-2-7b \
|
|
--use_slow_tokenizer \
|
|
--dataset_name HuggingFaceH4/ultrafeedback_binarized \
|
|
--max_seq_length 1024 \
|
|
--preprocessing_num_workers 128 \
|
|
--per_device_train_batch_size $BATCH_SIZE_PER_GPU \
|
|
--gradient_accumulation_steps $GRADIENT_ACC_STEPS \
|
|
--learning_rate 1e-4 \
|
|
--lr_scheduler_type linear \
|
|
--warmup_ratio 0.03 \
|
|
--weight_decay 0. \
|
|
--num_train_epochs 5 \
|
|
--output_dir output/tulu_v2_dpo_qlora/ \
|
|
--with_tracking \
|
|
--report_to tensorboard \
|
|
--logging_steps 1 &&
|
|
|
|
python open_instruct/merge_lora.py \
|
|
--base_model_name_or_path allenai/tulu-2-7b \
|
|
--lora_model_name_or_path output/tulu_v2_dpo_qlora/ \
|
|
--output_dir output/tulu_v2_dpo_qlora_merged/ \
|
|
--qlora \
|
|
--save_tokenizer
|