Skip to main content
Combine supervised fine-tuning and reinforcement learning in a single recipe. Train on demonstrations first to bootstrap output format, then continue with RL to improve quality.

​
Why two stages?

SFT teaches the model to imitate a curated set of demonstrations — picking up format, structure, and style from labeled examples. It generalizes from them, but remains anchored to their distribution: the model learns how good responses look, not necessarily how to reason toward them. RL builds on this foundation, optimizing behavior through feedback signals rather than imitation. Guided by a reward function, it can surface behaviors and strategies that no demonstration explicitly encoded — though the quality of what it discovers is ultimately bounded by the quality of the signal it’s optimizing for. Without SFT, RL must spend samples learning format basics — JSON structure, required fields, response conventions — rather than optimizing for quality. These constraints are far cheaper to teach through direct imitation than through reward signals, which are poorly suited to enforce rigid structural rules. SFT bootstraps format compliance efficiently, giving RL a well-initialized starting point. From there, RL can focus its signal on what matters: refining quality beyond what the demonstrations captured.

​
The two-stage recipe

This is a complete @recipe_main that orchestrates both SFT and RL in sequence. This example uses GSPO for the RL stage, but other algorithms (GRPO, ENVGRPO, ENVGSPO) work the same way: 
from adaptive_harmony.runtime import InputConfig, recipe_main
from adaptive_harmony.runtime.context import RecipeContext
from adaptive_harmony.common.sft import SFT
from adaptive_harmony.common.gspo import GSPO
from adaptive_harmony.parameters import Model, Dataset, dataset_kinds
from adaptive_harmony.graders import BinaryJudgeGrader
from adaptive_harmony.graders.combined_grader import CombinedGrader
from adaptive_harmony.metric_logger import get_prod_logger


class SFTtoGSPOConfig(InputConfig):
    """Configuration for SFT warm-up followed by GSPO training."""
    
    sft_dataset: Dataset[dataset_kinds.Completion]  # Demonstrations with expected outputs
    gspo_dataset: Dataset[dataset_kinds.Prompt]  # Prompts for RL (no ground truth needed)
    model_to_train: Model
    
    # SFT hyperparameters
    sft_learning_rate: float = 1e-5
    sft_epochs: int = 1
    sft_batch_size: int = 512
    
    # GSPO hyperparameters
    gspo_learning_rate: float = 5e-7
    gspo_max_steps: int = 1000
    gspo_batch_size: int = 256


@recipe_main
async def sft_to_gspo(config: SFTtoGSPOConfig, ctx: RecipeContext):
    """Run SFT warm-up, then continue with GSPO for quality improvement."""

    # --- Model setup ---
    # Both stages train the same model sequentially: SFT initializes it,
    # then GSPO continues from where SFT left off. We use spawn_train()
    # because we need gradient updates (spawn_inference is read-only).
    model_builder = await config.model_to_train.to_builder(ctx, tp=1, kv_cache_len=131072)
    model = await model_builder.spawn_train(
        "sft_then_gspo",
        max_batch_size=10_000,
    )

    logger = get_prod_logger(ctx=ctx)

    # === STAGE 1: SFT warm-up ===
    # SFT uses a Completion dataset — each sample contains a full conversation
    # with the expected assistant response. The model learns to reproduce these
    # demonstrations via supervised cross-entropy loss.
    sft_dataset = await config.sft_dataset.load(ctx)

    sft_trainer = SFT(
        dataset=sft_dataset,
        model=model,
        logger=logger,
        lr=config.sft_learning_rate,
        epochs=config.sft_epochs,           # Usually 1-3; more risks overfitting
        samples_per_batch=config.sft_batch_size,
        max_grad_norm=1.0,
        weight_decay=0.01,
    )
    await sft_trainer.run()
    # At this point, `model` has been updated in-place with SFT weights.
    # We continue training the same model instance in the next stage.

    # === STAGE 2: GSPO for quality improvement ===
    # GSPO uses a Prompt dataset — each sample is a prompt only, no expected
    # output. The algorithm generates completions, grades them, and learns
    # from the reward signal.
    gspo_dataset = await config.gspo_dataset.load(ctx)

    # Spawn a separate model for judging. This uses spawn_inference() (not
    # spawn_train) because we only need forward passes — no gradients.
    judge_model_builder = await config.model_to_train.to_builder(ctx, tp=1, kv_cache_len=131072)
    judge_model = await judge_model_builder.spawn_inference("judge")

    # Set up a grader to score completions. BinaryJudgeGrader sends each
    # completion to the judge LLM, which scores it PASS (1.0) or FAIL (0.0).
    grader = BinaryJudgeGrader(
        grader_key="output-quality",
        criteria="Is the output high-quality, well-formed, and correct?",
        model=judge_model,
    )

    # CombinedGrader aggregates one or more graders into a single score.
    # Even with one grader, the training algorithms expect a CombinedGrader.
    combined_grader = CombinedGrader(
        grader_key="combined",
        graders=[grader],
    )

    gspo_trainer = GSPO(
        dataset=gspo_dataset,
        model=model,                        # Same model instance, now SFT-warm-started
        grader=combined_grader,
        logger=logger,
        lr=config.gspo_learning_rate,       # Typically lower than SFT LR
        samples_per_batch=config.gspo_batch_size,
        samples_per_mini_batch=64,
        max_grad_norm=1.0,
        clip_range=0.2,                    
        kl_beta=0.1,                      
        max_num_gspo_steps=config.gspo_max_steps,
        completions_per_sample=4,           # Generate 4 completions per prompt for ranking
    )
    await gspo_trainer.run()

    # Save the final model — includes both SFT and GSPO training
    await model.save(model_name="sft-then-gspo-final", ctx=ctx)

​
SFT vs. RL: Complementary objectives

SFT (Supervised Fine-Tuning)
  • Learns to imitate labeled demonstrations — anchored to the distribution of examples
  • Uses supervised loss (typically cross-entropy)
  • Converges quickly: few epochs on a small demonstration set (early stopping avoids overfitting)
  • Best for: bootstrapping output format, structure, or style
RL (e.g., GSPO, GRPO, PPO)
  • Learns from feedback signals — grader scores, preference pairs, or other reward signals depending on the algorithm
  • Can discover behaviors beyond the demonstrations
  • Step-based; runs for many steps driven by online rollout generation, not dataset size
  • Best for: optimizing quality, refining output, and improving beyond what demonstrations capture
Together, they form a complete pipeline: imitate first, then improve.

​
When to skip SFT

  • Base model already formats correctly: If the vast majority of generations produce valid output, skip SFT and start with RL directly.
  • No demonstration data available: SFT requires labeled examples. If you only have unlabeled prompts, start with RL directly.
  • Output is simple: Plain text, single categories, or yes/no don’t need SFT. Complex, multi-field structured output benefits most.
  • Graders are lenient on format: If graders score semantically rather than penalizing parsing errors, malformed output is less costly.

​
Key takeaways

  1. SFT imitates, RL improves — SFT is anchored to its demonstrations; RL optimizes beyond them through reward signals
  2. SFT is efficient — A small demonstration set and few epochs converge quickly; stop early to avoid overfitting
  3. Dataset format depends on the algorithm — SFT needs full conversations with expected outputs; GSPO and GRPO use prompt-only datasets; DPO uses preference pairs
  4. Consider skipping SFT — If the base model already formats correctly, has no available demonstrations, or produces simple outputs, start directly with RL