Reinforcement Learning from Human Feedback (RLHF) examples: Direct Preference Optimization (DPO)

[danilopeixoto]

Introduce one Reinforcement Learning from Human Feedback (RLHF) example, such as Direct Preference Optimization (DPO) method.

Paper

Direct Preference Optimization: Your Language Model is Secretly a Reward Model

Notes

Direct Preference Optimization (DPO): A Simplified Explanation by João Lages

Implementation examples

• huggingface/trl: TRL - Transformer Reinforcement Learning
• eric-mitchell/direct-preference-optimization: Direct Preference Optimization

Possible MLX implementation

Policy and reference log probabilities:

def get_batched_logps(model, inputs, targets):
    logits, _ = model(inputs)
    logits = logits.astype(mx.float32)

    loss_mask = targets != 0
    per_token_logps = mx.take_along_axis(nn.log_softmax(logits), targets[..., None], axis=2).squeeze(2)

    return tuple((per_token_logps * loss_mask).sum(-1).split(2))

Loss:

def dpo_loss(model, beta, label_smoothing, reference_chosen_logps, reference_rejected_logps, inputs, targets):
    chosen_logps, rejected_logps = get_batched_logps(model, inputs, targets)

    pi_logratios = chosen_logps - rejected_logps
    reference_logratios = reference_chosen_logps - reference_rejected_logps

    logits = pi_logratios - reference_logratios
    losses = -nn.log_sigmoid(beta * logits) * (1.0 - label_smoothing) - nn.log_sigmoid(-beta * logits) * label_smoothing

    chosen_rewards = beta * (chosen_logps - reference_chosen_logps)
    rejected_rewards = beta * (rejected_logps - reference_rejected_logps)
    reward_accuracies = (chosen_rewards > rejected_rewards).astype(mx.float32)
    reward_margins = chosen_rewards - rejected_rewards

    ntoks = (inputs != 0).sum()

    return (
        losses.mean(),
        chosen_rewards.mean(),
        rejected_rewards.mean(),
        reward_accuracies.mean(),
        reward_margins.mean(),
        ntoks,
    )

Beta: The temperature parameter for the DPO loss is typically set in the range of 0.1 to 0.5. The reference model is ignored when beta equals 0.

Label smoothing: This parameter represents the conservativeness for DPO loss, assuming that preferences are noisy and can be flipped with a probability of label_smoothing.

Note label_smoothing > 0 defines the Conservative DPO loss.

[awni]

@danilopeixoto I've been thinking about having this in MLX LM recently. Any interest in sending a PR?

It might make to do it after we have a more manageable config (#503) but that should be landed soon!

[awni]

To be more concrete, I'm envisioning you just set the loss in the config. e.g. cross_entropy or dpo

[ivanfioravanti]

This would be an awesome addition to mlx_examples! 🔥

[N8python]

I'm very very excited for this! Don't have the technical expertise to implement the DPO directly but would love to help in other ways (config, code cleanup) if neccessary!

[lin72h]

That makes MLX really useful for production not just a research tool!

[kishoretvk]

+500 waiting for this

[developerlin]

Wait for this, when will the DPO training be supported?

[Goekdeniz-Guelmez]

#1233 #1210 #1209