Cookbook: Distillation

Run on-policy distillation on Fireworks using cookbook recipes that combine frozen teacher scoring, student sampling, and reward-shaped loss.

​

What this is

The cookbook’s training.recipes.distillation_loop trains one student from its own rollouts while a frozen teacher scores those same responses.

Use it when you want recipe-managed trainer provisioning, student sampling, teacher scoring, checkpointing, and cleanup for distillation experiments.

​

Recipes

Recipe	Teacher signal
distillation_loop	Sampled-token OPD: teacher logprob minus sampling logprob on the student’s own completions.
External SDFT cookbook	Top-K self-distillation on serverless, with privileged answer context.

SDFT is not part of this cookbook. It is maintained in the Fireworks Tinker-Cookbook repository; follow the SDFT guide and use a separate Python environment. It is a different objective from distillation_loop, not a flag on it. See the distillation skill for teacher setup and algorithm details.

​

Minimal example

from training.recipes.distillation_loop import Config, main
from training.utils import DeployConfig, TrainerConfig

cfg = Config(
    log_path="./distillation_logs",
    base_model="accounts/fireworks/models/qwen3-8b",
    teacher_model="accounts/fireworks/models/qwen3-32b",
    dataset="/path/to/prompts.jsonl",
    trainer=TrainerConfig(
        training_shape_id="accounts/fireworks/trainingShapes/qwen3-8b-128k",
    ),
    deployment=DeployConfig(tokenizer_model="Qwen/Qwen3-8B"),
    max_rows=100,
    epochs=1,
)

main(cfg)

If teacher_model is a base model resource, the recipe creates a frozen teacher deployment for scoring. If it is already an inference model or deployment resource, the recipe uses it directly.

​

Multi-teacher runs

The current OPD recipe uses one frozen teacher. Fork its loop to route prompts to multiple teachers.

​

Dataset contract

Rows are JSONL objects. The only required field is messages, the student-visible OpenAI-style chat prompt.

Optional fields:

Field	Use
teacher_messages	Teacher-side prompt for privileged-context scoring. If omitted, the teacher scores under messages.
expected_answer	Optional metadata for eval callbacks and smoke checks.

Student and teacher token IDs must use a compatible tokenizer and vocabulary. Prefer teachers from the same model family, and use a matching tokenizer for DeployConfig.tokenizer_model.

​

Examples

The cookbook includes distillation examples under training/examples/distillation:

Example	Path	Use
Privileged-context OPD	gsm8k_privileged	Student sees the problem; teacher can see privileged solution context.
Serverless privileged distillation	serverless_k3_privileged	LoRA student trained from teacher scores on a privileged prompt.

Run from the cookbook repository:

cd training
FIREWORKS_API_KEY=... \
python -m training.examples.distillation.gsm8k_privileged.train_gsm8k_privileged --help

​

Next steps

• Cookbook Reference - config classes and common recipe fields
• Loss Functions - built-in and custom Training API losses
• Weight sync - how updated weights reach serving deployments