Direct Preference Optimization (DPO)

Overview

Direct Preference Optimization (DPO) is an alignment technique that fine-tunes foundation models using paired comparison data to align model outputs with human preferences. Unlike reinforcement learning methods, DPO directly optimizes model behavior based on human feedback about which responses are more desirable, offering a more stable and scalable approach.

Why use DPO

Foundation models may generate outputs that are factually correct but fail to align with specific user needs, organizational values, or safety requirements. DPO addresses this by enabling you to:

How DPO works

DPO uses paired examples where human evaluators indicate which of two possible responses is preferred. The model learns to maximize the likelihood of generating preferred responses while minimizing undesired ones.

When to use DPO

Use DPO in the following scenarios:

Supported models and techniques

DPO supports both full-parameter fine-tuning and LoRA (Low-Rank Adaptation):

Training approaches

For most use cases, the LoRA approach provides sufficient adaptation capability with significantly improved efficiency.

Data format

DPO training data follows the same format as SFT, except the last assistant turn must contain preference pairs with preferred and non-preferred labels.

Basic structure

The final assistant turn uses a candidates array instead of content:

{
  "role": "assistant",
  "candidates": [
    {
      "content": [
        {
          "text": "This is the preferred response."
        }
      ],
      "preferenceLabel": "preferred"
    },
    {
      "content": [
        {
          "text": "This is the non-preferred response."
        }
      ],
      "preferenceLabel": "non-preferred"
    }
  ]
}

Complete text example

{
  "schemaVersion": "bedrock-conversation-2024",
  "system": [
    {
      "text": "You are a helpful assistant."
    }
  ],
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "text": "What is the capital of France?"
        }
      ]
    },
    {
      "role": "assistant",
      "content": [
        {
          "text": "The capital of France is Paris."
        }
      ]
    },
    {
      "role": "user",
      "content": [
        {
          "text": "Tell me more about it."
        }
      ]
    },
    {
      "role": "assistant",
      "candidates": [
        {
          "content": [
            {
              "text": "Paris is the capital and largest city of France, known for the Eiffel Tower, world-class museums like the Louvre, and its rich cultural heritage."
            }
          ],
          "preferenceLabel": "preferred"
        },
        {
          "content": [
            {
              "text": "Paris is a city in France."
            }
          ],
          "preferenceLabel": "non-preferred"
        }
      ]
    }
  ]
}

Example with images

{
  "schemaVersion": "bedrock-conversation-2024",
  "system": [
    {
      "text": "You are a helpful assistant."
    }
  ],
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "text": "Describe this image."
        },
        {
          "image": {
            "format": "jpeg",
            "source": {
              "s3Location": {
                "uri": "s3://your-bucket/your-path/image.jpg",
                "bucketOwner": "your-aws-account-id"
              }
            }
          }
        }
      ]
    },
    {
      "role": "assistant",
      "candidates": [
        {
          "content": [
            {
              "text": "The image shows a detailed description with relevant context and observations."
            }
          ],
          "preferenceLabel": "preferred"
        },
        {
          "content": [
            {
              "text": "This is a picture."
            }
          ],
          "preferenceLabel": "non-preferred"
        }
      ]
    }
  ]
}

Dataset requirements

Uploading data

aws s3 cp /path/to/training-data/ s3://your-bucket/train/ --recursive
aws s3 cp /path/to/validation-data/ s3://your-bucket/val/ --recursive

Recipe configuration

General run configuration

run:
  name: "my-dpo-run"
  model_type: "amazon.nova-lite-v1:0:300k"
  model_name_or_path: "nova-lite/prod"
  replicas: 4

Training configuration

training_config:
  max_length: 16384
  global_batch_size: 32

  trainer:
    max_epochs: 3

  model:
    hidden_dropout: 0.0
    attention_dropout: 0.0
    ffn_dropout: 0.0

Optimizer configuration

model:
  optim:
    lr: 1e-5
    name: distributed_fused_adam
    adam_w_mode: true
    eps: 1e-08
    weight_decay: 0.0
    betas:
      - 0.9
      - 0.999
    sched:
      warmup_steps: 10
      constant_steps: 0
      min_lr: 1e-6

DPO-specific configuration

model:
  dpo_cfg:
    beta: 0.1

Recommendation: Start with beta: 0.1 and adjust downward if preference learning seems insufficient.

LoRA PEFT configuration

model:
  peft:
    peft_scheme: "lora"
    lora_tuning:
      loraplus_lr_ratio: 64.0
      alpha: 32
      adapter_dropout: 0.01

Starting a training job

Container image

708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest

Example code

from sagemaker.pytorch import PyTorch
from sagemaker.inputs import TrainingInput

instance_type = "ml.p5.48xlarge"
instance_count = 4

image_uri = "708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest"

recipe_overrides = {
    "training_config": {
        "trainer": {"max_epochs": 2},
        "model": {
            "dpo_cfg": {"beta": 0.1},
            "peft": {
                "peft_scheme": "lora",
                "lora_tuning": {
                    "loraplus_lr_ratio": 64.0,
                    "alpha": 32,
                    "adapter_dropout": 0.01,
                },
            },
        },
    },
}

estimator = PyTorch(
    output_path=f"s3://{bucket_name}/{job_name}",
    base_job_name=job_name,
    role=role,
    instance_count=instance_count,
    instance_type=instance_type,
    training_recipe="fine-tuning/nova/nova_lite_p5_gpu_lora_dpo",
    recipe_overrides=recipe_overrides,
    max_run=18000,
    sagemaker_session=sagemaker_session,
    image_uri=image_uri,
    disable_profiler=True,
    debugger_hook_config=False,
)

train_input = TrainingInput(
    s3_data=train_dataset_s3_path,
    distribution="FullyReplicated",
    s3_data_type="Converse",
)

val_input = TrainingInput(
    s3_data=val_dataset_s3_path,
    distribution="FullyReplicated",
    s3_data_type="Converse",
)

estimator.fit(inputs={"train": train_input, "validation": val_input}, wait=True)

Deploying the model

After training completes, deploy the customized model to Amazon Bedrock using the Custom Model Import functionality. The model supports both provisioned throughput and on-demand inference. LoRA-trained models support on-demand inference.

For deployment instructions, see Deploying customized models.

Limitations