AdamG012

/

chat-opt-1.3b-rlhf-actor-deepspeed

like 5

I have noticed that you used 4 datasets in step 1 and 2 but in step 3 only Dahoas/rm-static why ? is there a reason?
Does the step 3 depends on the step 1 and 2 and no need to attach the whole datasets in rlhf step 3 ?

actor cmd
--data_path Dahoas/rm-static Dahoas/full-hh-rlhf Dahoas/synthetic-instruct-gptj-pairwise yitingxie/rlhf-reward-datasets --data_split 2,4,4 --model_name_or_path facebook/opt-1.3b --per_device_train_batch_size 8 --per_device_eval_batch_size 8 --max_seq_len 512 --learning_rate 9.65e-6 --weight_decay 0. --num_train_epochs 16 --gradient_accumulation_steps 1 --lr_scheduler_type cosine --num_warmup_steps 0 --seed 1234 --zero_stage 2 --deepspeed --output_dir ./output

reward cmd
--data_path Dahoas/rm-static Dahoas/full-hh-rlhf Dahoas/synthetic-instruct-gptj-pairwise yitingxie/rlhf-reward-datasets --data_split 2,4,4 --model_name_or_path facebook/opt-350m --num_padding_at_beginning 1 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --max_seq_len 512 --learning_rate 5e-5 --weight_decay 0.1 --num_train_epochs 1 --disable_dropout --gradient_accumulation_steps 1 --lr_scheduler_type cosine --num_warmup_steps 0 --seed 1234 --zero_stage 0 --deepspeed --output_dir ./output

rlhf actor
--data_path Dahoas/rm-static --data_split 2,4,4 --actor_model_name_or_path ../step1_supervised_finetuning/output/ --critic_model_name_or_path ../step2_reward_model_finetuning/output/ --num_padding_at_beginning 1 --per_device_train_batch_size 4 --per_device_mini_train_batch_size 4 --generation_batch_numbers 1 --ppo_epochs 1 --max_answer_seq_len 256 --max_prompt_seq_len 256 --actor_learning_rate 9.65e-6 --critic_learning_rate 5e-6 --num_train_epochs 1 --lr_scheduler_type cosine --gradient_accumulation_steps 1 --disable_actor_dropout --num_warmup_steps 100 --deepspeed --seed 1234 --enable_hybrid_engine --actor_zero_stage 2 --critic_zero_stage 2 --enable_ema --output_dir ./output

rlhf critic
--data_path Dahoas/rm-static --data_split 2,4,4 --actor_model_name_or_path ../step1_supervised_finetuning/output/ --critic_model_name_or_path ../step2_reward_model_finetuning/output/ --num_padding_at_beginning 1 --per_device_train_batch_size 4 --per_device_mini_train_batch_size 4 --generation_batch_numbers 1 --ppo_epochs 1 --max_answer_seq_len 256 --max_prompt_seq_len 256 --actor_learning_rate 9.65e-6 --critic_learning_rate 5e-6 --num_train_epochs 1 --lr_scheduler_type cosine --gradient_accumulation_steps 1 --disable_actor_dropout --num_warmup_steps 100 --deepspeed --seed 1234 --enable_hybrid_engine --actor_zero_stage 2 --critic_zero_stage 2 --enable_ema --output_dir ./output