Efficient fine-tuning in RLHF

Reinforcement Learning from Human Feedback (RLHF)

Mina Parham

AI Engineer

Parameter-efficient fine-tuning

Fine-tuning a full model

Illustration representing a neural network.

Parameter-efficient fine-tuning

Fine-tuning with PEFT

Illustration representing a neural network with most of the parameters frozen.

LoRA: adjusts only a few layers
Quantization: lowers data type precision

Step1: load your active model in 8-bit precision

from peft import prepare_model_for_int8_training


pretrained_model = AutoModelForCausalLM.from_pretrained(
                                                        model_name, 
                                                        load_in_8bit=True
                                                       )

pretrained_model_8bit = prepare_model_for_int8_training(pretrained_model)

Step 2: add extra trainable adapters using peft

from peft import LoraConfig, get_peft_model


config = LoraConfig(

    r=32, # Rank of the low-rank matrices

    lora_alpha=32, # Scaling factor for the LoRA updates

    lora_dropout=0.1, # Dropout rate for LoRA layers

    bias="lora_only"# Only update bias terms for LoRA layers, others remain frozen

)


lora_model = get_peft_model(pretrained_model_8bit, config)
model = AutoModelForCausalLMWithValueHead.from_pretrained(lora_model)

Step 3: use one model for reference and active logits

ppo_trainer = PPOTrainer(
    config, # The config we just defined
    model, # Our PPO model
    ref_model=None, 
    tokenizer=tokenizer, 
    dataset=dataset, 
    data_collator=collator, 
    optimizer=optimizer
)

Let's practice!

Reinforcement Learning from Human Feedback (RLHF)