SDPO ↗

Self-Distillation Policy Optimization (SDPO) was introduced in Reinforcement Learning via Self-Distillation by Jonas Hübotter, Frederike Lübeck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, and Andreas Krause.

Large language models are increasingly post-trained with reinforcement learning in verifiable domains such as code and math. Yet, current methods for reinforcement learning with verifiable rewards (RLVR) learn only from a scalar outcome reward per attempt, creating a severe credit-assignment bottleneck. Many verifiable environments actually provide rich textual feedback, such as runtime errors or judge evaluations, that explain why an attempt failed. We formalize this setting as reinforcement learning with rich feedback and introduce Self-Distillation Policy Optimization (SDPO), which converts tokenized feedback into a dense learning signal without any external teacher or explicit reward model. SDPO treats the current model conditioned on feedback as a self-teacher and distills its feedback-informed next-token predictions back into the policy. In this way, SDPO leverages the model’s ability to retrospectively identify its own mistakes in-context. Across scientific reasoning, tool use, and competitive programming on LiveCodeBench v6, SDPO improves sample efficiency and final accuracy over strong RLVR baselines. Notably, SDPO also outperforms baselines in standard RLVR environments that only return scalar feedback by using successful rollouts as implicit feedback for failed attempts. Finally, applying SDPO to individual questions at test time accelerates discovery on difficult binary-reward tasks, achieving the same discovery probability as best-of-k sampling or multi-turn conversations with 3x fewer attempts.

The SDPO trainer is built on TRL’s experimental shared self-distillation stack. It keeps the online rollout-and-reward training flow, then builds a teacher-conditioned view of the same completions from successful rollouts and optional environment feedback.

In the current TRL implementation:

• the default SDPO policy loss mode is distillation_only
• hybrid mode is also available to combine the base policy loss with the self-distillation loss
• supported teacher regularization modes are ema and none
• distillation_topk is only valid when full_logit_distillation=True
• when full_logit_distillation=False, SDPO uses token-level reverse KL and requires distillation_alpha=1.0
• environment feedback can be injected into teacher reprompts when the dataset exposes a privileged_context column

Expected dataset columns#

Each example must provide:

• prompt: the student-facing prompt
• privileged_context: optional privileged text, such as environment feedback, used when include_environment_feedback=True

Usage#

from datasets import Dataset

from trl.experimental.sdpo import SDPOConfig, SDPOTrainer

dataset = Dataset.from_dict(
    {
        "prompt": [[{"role": "user", "content": "Solve 2+2."}]],
        "privileged_context": ["Your earlier answer used the wrong format."],
    }
)

training_args = SDPOConfig(
    output_dir="sdpo-model",
    distillation_topk=100,                 # Top-K logit distillation approximation
    full_logit_distillation=True,          # Required for top-K; enables non-reverse divergences
    include_environment_feedback=True,     # Use dataset privileged_context for teacher reprompts
)

trainer = SDPOTrainer(
    model="Qwen/Qwen2.5-1.5B-Instruct",
    reward_funcs=reward_func,
    args=training_args,
    train_dataset=dataset,
)
trainer.train()

SDPO always requires a prompt column. To use environment feedback, also include a privileged_context column and set include_environment_feedback=True. SDPO will use successful rollouts and, when enabled, that text to build teacher reprompts for self-distillation.

Callbacks#

The trainer emits a small set of callback hooks that are useful for debugging, observability, and tests. These hooks are intended as practical integration points for experimental self-distillation workflows.

Shared self-distillation hooks:

• on_self_distillation_batch_prepared: fired when a self-distillation batch is ready. The payload includes prompt_ids, completion_ids, and old_per_token_logps when importance-sampling clipping inputs are available.
• on_generation_batch_built: fired when a new buffered generation batch is created. The payload includes generate_every and steps_per_generation.

SDPO-specific hook:

• on_teacher_context_built: fired after SDPO constructs the teacher-conditioned inputs. The payload includes teacher_input_ids, teacher_attention_mask, completion_mask, and self_distillation_mask.

Example script#

Use trl/experimental/sdpo/sdpo.py to launch SDPO training from the command line. The script supports verifiable math rewards, environment feedback via --feedback_column, and PEFT/LoRA via the standard ModelConfig flags.

python trl/experimental/sdpo/sdpo.py \
    --model_name_or_path Qwen/Qwen2.5-Math-1.5B-Instruct \
    --dataset_name openai/gsm8k \
    --dataset_config main \
    --output_dir outputs/sdpo-qwen35-2b-gsm8k \
    --learning_rate 5e-5 \
    --dtype bfloat16 \
    --bf16 true \
    --max_completion_length 128 \
    --use_peft \
    --lora_target_modules q_proj k_proj v_proj o_proj gate_proj up_proj down_proj \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 2 \
    --num_generations 8 \
    --generation_batch_size 32 \
    --distillation_alpha 1.0 \
    --full_logit_distillation false \
    --sdpo_policy_loss_mode hybrid \
    --report_to none \
    --eval_strategy steps \
    --eval_steps 1000 \
    --save_strategy no

SDPOConfig[[trl.experimental.sdpo.SDPOConfig]]#

trl.experimental.sdpo.SDPOConfig[[trl.experimental.sdpo.SDPOConfig]]#

Source

Configuration class for the SDPOTrainer.

This class extends experimental.self_distillation.SelfDistillationConfig with the online teacher-construction parameters used by Self-Distillation Policy Optimization (SDPO).

SDPOTrainer[[trl.experimental.sdpo.SDPOTrainer]]#

trl.experimental.sdpo.SDPOTrainer[[trl.experimental.sdpo.SDPOTrainer]]#

Source

Trainer for Self-Distillation Policy Optimization (SDPO).

SDPO augments on-policy optimization with self-distillation from the model’s own high-reward trajectories. It converts tokenized feedback into a dense learning signal without any external teacher or explicit reward model. SDPO treats the current model conditioned on feedback as a self-teacher and distills its feedback-informed next-token predictions back into the policy.

traintrl.experimental.sdpo.SDPOTrainer.trainhttps://github.com/huggingface/trl/blob/v1.5.1/transformers/trainer.py#L1325[{“name”: “resume_from_checkpoint”, “val”: “: str | bool | None = None”}, {“name”: “trial”, “val”: “: optuna.Trial | dict[str, Any] | None = None”}, {“name”: “ignore_keys_for_eval”, “val”: “: list[str] | None = None”}]- resume_from_checkpoint (str or bool, optional) – If a str, local path to a saved checkpoint as saved by a previous instance of Trainer. If a bool and equals True, load the last checkpoint in args.output_dir as saved by a previous instance of Trainer. If present, training will resume from the model/optimizer/scheduler states loaded here.

• trial (optuna.Trial or dict[str, Any], optional) – The trial run or the hyperparameter dictionary for hyperparameter search.
• ignore_keys_for_eval (list[str], optional) – A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions for evaluation during the training.0~trainer_utils.TrainOutputObject containing the global step count, training loss, and metrics.

Main training entry point.

Parameters:

resume_from_checkpoint (str or bool, optional) : If a str, local path to a saved checkpoint as saved by a previous instance of Trainer. If a bool and equals True, load the last checkpoint in args.output_dir as saved by a previous instance of Trainer. If present, training will resume from the model/optimizer/scheduler states loaded here.

trial (optuna.Trial or dict[str, Any], optional) : The trial run or the hyperparameter dictionary for hyperparameter search.

ignore_keys_for_eval (list[str], optional) : A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions for evaluation during the training.

Returns:

~trainer_utils.TrainOutput

Object containing the global step count, training loss, and metrics.

save_model[[trl.experimental.sdpo.SDPOTrainer.save_model]]#

Source

Will save the model, so you can reload it using from_pretrained().

Will only save from the main process.

push_to_hub[[trl.experimental.sdpo.SDPOTrainer.push_to_hub]]#

Source

Upload self.model and self.processing_class to the 🤗 model hub on the repo self.args.hub_model_id.

Parameters:

commit_message (str, optional, defaults to "End of training") : Message to commit while pushing.

blocking (bool, optional, defaults to True) : Whether the function should return only when the git push has finished.

token (str, optional, defaults to None) : Token with write permission to overwrite Trainer’s original args.

revision (str, optional) : The git revision to commit from. Defaults to the head of the “main” branch.

kwargs (dict[str, Any], optional) : Additional keyword arguments passed along to ~Trainer.create_model_card.

Returns:

The URL of the repository where the model was pushed if blocking=False, or a Future object tracking the progress of the commit if blocking=True.