Skip to main content

​
What this is

This guide walks through DPO (Direct Preference Optimization) training using the cookbook. DPO learns from preference pairs (chosen vs. rejected responses) without a separate reward model.

​
How DPO differs from GRPO

DPOGRPO
Trainer jobs1 for LoRA, 2 for full-parameter (policy + frozen reference)1-2 trainers plus an inference deployment, depending on reference needs
DataPreference pairs (chosen/rejected)Prompts + reward function
Reference logprobsCached once at initializationComputed every step
Loss-log(sigmoid(beta * margin))Advantage-weighted policy gradient + KL

​
Architecture

​
Using the recipe

from training.recipes.dpo_loop import Config, main
from training.utils import TrainerConfig, WandBConfig

cfg = Config(
    log_path="./dpo_logs",
    base_model="accounts/fireworks/models/qwen3-8b",
    dataset="/path/to/preference_data.jsonl",
    tokenizer_model="Qwen/Qwen3-8B",
    beta=0.1,
    epochs=1,
    batch_size=4,
    max_seq_len=4096,
    trainer=TrainerConfig(
        training_shape_id="accounts/fireworks/trainingShapes/qwen3-8b-128k-h200",
        reference_training_shape_id="accounts/fireworks/trainingShapes/qwen3-8b-128k-h200-forward",
    ),
    wandb=WandBConfig(entity="my-team", project="dpo-experiment"),
)

main(cfg)

​
Dataset format

DPO expects preference pairs. Supported formats: Format 1 — chosen/rejected messages: 
{
  "chosen": {"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "good response"}]},
  "rejected": {"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "bad response"}]}
}
Format 2 — input/output split: 
{
  "input": {"messages": [{"role": "user", "content": "..."}]},
  "preferred_output": [{"role": "assistant", "content": "good"}],
  "non_preferred_output": [{"role": "assistant", "content": "bad"}]
}

​
Step-by-step (API-level)

​
Provision trainers with build_service_client

DPO always needs reference logprobs. Full-parameter DPO uses a policy trainer and a forward-only reference trainer; LoRA DPO uses one policy trainer and the policy session’s shared base reference. Provisioning is owned by the SDK-managed service client — build_service_client resolves shapes, attaches or creates the trainer(s), and decides the reference strategy for you:
  • LoRA (lora_rank > 0) with no reference_training_shape_id → create_reference_client reuses the policy session (no second trainer).
  • Full-parameter, or an explicit reference_training_shape_id → a separate forward-only reference trainer is provisioned and its lifecycle is owned by the service client.
import os
from training.utils import TrainerConfig, build_service_client

api_key = os.environ["FIREWORKS_API_KEY"]
base_url = os.environ.get("FIREWORKS_BASE_URL", "https://api.fireworks.ai")
base_model = "accounts/fireworks/models/qwen3-8b"

service = build_service_client(
    api_key=api_key,
    base_url=base_url,
    additional_headers=None,
    base_model=base_model,
    tokenizer_model="Qwen/Qwen3-8B",
    lora_rank=0,
    max_context_length=None,
    learning_rate=1e-5,
    trainer=TrainerConfig(
        training_shape_id="accounts/fireworks/trainingShapes/qwen3-8b-128k-h200",
        reference_training_shape_id="accounts/fireworks/trainingShapes/qwen3-8b-128k-h200-forward",
    ),
    # deployment=None  → trainer-only provisioning (DPO has no rollouts)
    cleanup_trainer_on_close=True,  # delete SDK-managed trainers on service.close()
)

policy_client = service.create_training_client(base_model, lora_rank=0)
reference_client = service.create_reference_client(base_model, lora_rank=0)

# ... training loop ...
# service.close()  # tears down the trainers it created
The cookbook recipes wrap these clients in ReconnectableClient.from_training_client(...) for blocking semantics; for a raw API-level loop you can call policy_client / reference_client directly.

​
Cache reference logprobs

Reference logprobs are computed once at initialization and reused throughout training: 
ref_cache = {}
for i, (chosen_tokens, rejected_tokens, prompt_len) in enumerate(dataset):
    chosen_datum, rejected_datum = build_dpo_datums(
        chosen_tokens, rejected_tokens, prompt_len, max_seq_len=4096,
    )
    fwd = reference_client.forward([chosen_datum, rejected_datum], "cross_entropy")
    ref_cache[i] = {
        "ref_chosen": fwd.loss_fn_outputs[0]["logprobs"].data,
        "ref_rejected": fwd.loss_fn_outputs[1]["logprobs"].data,
        "chosen_tokens": chosen_tokens,
        "rejected_tokens": rejected_tokens,
        "prompt_len": prompt_len,
    }

​
DPO loss function

import torch
import torch.nn.functional as F

def make_dpo_loss_fn(ref_chosen_logprobs, ref_rejected_logprobs, beta=0.1):
    ref_chosen_t = torch.tensor(ref_chosen_logprobs, dtype=torch.float32)
    ref_rejected_t = torch.tensor(ref_rejected_logprobs, dtype=torch.float32)

    def loss_fn(data, logprobs_list):
        pi_chosen, pi_rejected = logprobs_list[0], logprobs_list[1]
        chosen_weights = torch.tensor(data[0].loss_fn_inputs["weights"].data, dtype=torch.float32)
        rejected_weights = torch.tensor(data[1].loss_fn_inputs["weights"].data, dtype=torch.float32)

        pi_chosen_sum = torch.dot(pi_chosen.float(), chosen_weights)
        pi_rejected_sum = torch.dot(pi_rejected.float(), rejected_weights)
        ref_chosen_sum = torch.dot(ref_chosen_t.float(), chosen_weights)
        ref_rejected_sum = torch.dot(ref_rejected_t.float(), rejected_weights)

        margin = (pi_chosen_sum - ref_chosen_sum) - (pi_rejected_sum - ref_rejected_sum)
        dpo_loss = -F.logsigmoid(beta * margin)

        with torch.no_grad():
            accuracy = 1.0 if margin.item() > 0 else 0.0

        return dpo_loss, {"dpo_loss": dpo_loss.item(), "margin": margin.item(), "accuracy": accuracy}

    return loss_fn

​
Training loop

step = 0
accum_count = 0
grad_accum = 4

for idx in ref_cache:
    cached = ref_cache[idx]
    chosen_datum, rejected_datum = build_dpo_datums(
        cached["chosen_tokens"], cached["rejected_tokens"],
        cached["prompt_len"], max_seq_len=4096,
    )
    loss_fn = make_dpo_loss_fn(
        ref_chosen_logprobs=cached["ref_chosen"],
        ref_rejected_logprobs=cached["ref_rejected"],
        beta=0.1,
    )
    result = policy_client.forward_backward_custom([chosen_datum, rejected_datum], loss_fn)
    accum_count += 1

    if accum_count >= grad_accum:
        policy_client.optim_step(
            tinker.AdamParams(learning_rate=1e-5, beta1=0.9, beta2=0.999, eps=1e-8, weight_decay=0.01)
        )
        step += 1
        accum_count = 0
        print(f"Step {step}: {result.metrics}")

​
Operational guidance

  • Set trainer.training_shape_id when you need an explicit policy shape — otherwise supported recipes auto-select a validated policy shape.
  • Leave trainer.reference_training_shape_id unset unless you need a specific reference shape — full-parameter DPO auto-selects a forward-only reference shape; LoRA DPO uses a shared-session reference by default.
  • DPO does not provision a deployment — there are no rollout samples or deployment weight syncs in the recipe.
  • Keep a versioned reference cache tied to tokenizer + base model revision. If the base model changes, recompute reference logprobs.
  • Monitor margin statistics: increasing margins indicate the policy is learning preferences.
  • DCP checkpoints are disabled by default (dcp_save_interval=0). If you need to resume training from a checkpoint, set dcp_save_interval directly on dpo_loop.Config.

​
Common pitfalls

  • Mismatched formatting between chosen/rejected sequences corrupts preference signals — ensure identical prompt prefixes.
  • Stale reference cache: If you warm-start from a different model, cached reference logprobs are invalid.
  • ORPO (training.recipes.orpo_loop) — Odds Ratio Preference Optimization. Combines an SFT-style negative-log-likelihood term on the chosen response with a margin term on the odds ratio between chosen and rejected. Unlike DPO, ORPO does not require a reference trainer (no cached reference logprobs), so the recipe runs with a single trainer + dataset of preference pairs. See training.recipes.orpo_loop in the public cookbook repo for the full configuration.