Cookbook: SFT ↗

Documentation Index#

Fetch the complete documentation index at: https://docs.fireworks.ai/llms.txt Use this file to discover all available pages before exploring further.

Supervised fine-tuning via the cookbook’s sft_loop recipe.

What this is#

Supervised Fine-Tuning (SFT) trains the model to produce desired outputs by minimizing cross-entropy loss on (prompt, response) pairs. The cookbook’s sft_loop recipe handles data loading, tokenization, batching, gradient accumulation, and checkpointing automatically.

Using the recipe#

from training.recipes.sft_loop import Config, main
from training.utils import TrainerConfig

cfg = Config(
    log_path="./sft_logs",
    base_model="accounts/fireworks/models/qwen3-8b",
    dataset="/path/to/training_data.jsonl",
    tokenizer_model="Qwen/Qwen3-8B",
    max_seq_len=4096,
    epochs=1,
    batch_size=4,
    learning_rate=1e-5,
    trainer=TrainerConfig(
        training_shape_id="accounts/fireworks/trainingShapes/qwen3-8b-128k-h200",
    ),
)

main(cfg)

batch_size_samples is not supported in the V2 SFT CookbookTrainingConfig.

Passing batch_size_samples to the V2 config has no effect — the parameter is accepted without error but silently ignored, which can lead to unexpected step counts.

How batching works in V2: Steps are calculated as:

steps = (num_samples × num_epochs) / batch_size

where batch_size is set by the training shape and the recipe’s batch_size field — not by batch_size_samples.

Example: 10 samples × 5 epochs ÷ batch size of 10 = 5 steps, not 50.

To control training length, adjust epochs (and related recipe fields). Contact support for custom batch size configurations.

Dataset format#

SFT datasets use the standard messages format (JSONL with one example per line):

{"messages": [
  {"role": "user", "content": "What is the capital of France?"},
  {"role": "assistant", "content": "The capital of France is Paris."}
]}

Multi-turn conversations are supported:

{"messages": [
  {"role": "system", "content": "You are a helpful assistant."},
  {"role": "user", "content": "Hello"},
  {"role": "assistant", "content": "Hi! How can I help?"},
  {"role": "user", "content": "What is 2+2?"},
  {"role": "assistant", "content": "2+2 = 4"}
]}

The recipe automatically tokenizes conversations using the chat template, setting token weights to 0.0 for prompt tokens and 1.0 for response tokens.

Vision datasets#

The SFT recipe also supports vision-language model fine-tuning. Use multimodal content arrays with image_url objects in your JSONL, and specify a VLM training shape and tokenizer. See Vision Inputs for dataset format details and a full walkthrough.

Checkpointing and resume#

The current sft_loop recipe manages the trainer-side loop only. It does not create a deployment or run weight sync during training, but it does expose DCP checkpointing and resume controls:

from training.utils import TrainerConfig, WandBConfig

cfg = Config(
    log_path="./sft_logs",
    base_model="accounts/fireworks/models/qwen3-8b",
    dataset="/path/to/training_data.jsonl",
    tokenizer_model="Qwen/Qwen3-8B",
    max_seq_len=4096,
    epochs=1,
    batch_size=4,
    trainer=TrainerConfig(
        training_shape_id="accounts/fireworks/trainingShapes/qwen3-8b-128k-h200",
    ),
    dcp_save_interval=50,
    init_from_checkpoint="previous-job-id:step-100",  # optional
    wandb=WandBConfig(entity="my-team", project="sft-experiment"),
)

main(cfg)

Operational guidance#

• Set trainer.training_shape_id — cookbook trainer launches use training shapes.
• Only one trainer job needed — SFT does not require a reference trainer.
• The current recipe does not provision a deployment — use the API directly if you want deployment-side evaluation or weight sync during SFT.
• Use batch_size to control the number of examples per optimizer step.
• Gradient accumulation normalization defaults to None — the SFT loss is already normalized client-side, so adding server-side normalization would double-normalize gradients.
• Resume: The recipe uses checkpoint_utils.resolve_resume() to automatically restore from the last saved state on restart.
• DCP checkpoints are disabled by default (dcp_save_interval=0). If you need to resume training from a checkpoint, you must explicitly set dcp_save_interval to a positive value (e.g., dcp_save_interval=50).

Related guides#

• Vision Inputs — VLM fine-tuning with image and text data
• Cookbook DPO — preference optimization
• Cookbook RL (GRPO) — reinforcement learning recipes
• Cookbook Reference — all config classes and parameters
• Loss Functions — API-level SFT loss details