Supervised fine-tuning (SFT)

Supervised fine-tuning (method="sft_text") trains a model to imitate your labeled examples. It is the simplest way to adapt a base model to a task or style, and a common first pass before GRPO: teach the format with SFT, then reinforce correctness with RL. Veri runs SFT on TRL’s SFTTrainer. You provide a dataset of text or chat examples; Veri provisions the GPU, runs the loop, streams metrics, and uploads the checkpoint.

Dataset format

SFT accepts either plain text or conversational rows. For chat data, the trainer applies the model’s chat template automatically.

# Plain text (the `dataset_text_field`, default "text")
{"text": "The sky is blue."}

# Conversational — chat template applied automatically
{"messages": [{"role": "user", "content": "What color is the sky?"},
              {"role": "assistant", "content": "It is blue."}]}

Submitting a job

This mirrors TRL’s SFT quickstart (a small Qwen model on the Capybara chat dataset), run on Veri:

from veri_sdk import Client

client = Client()

dataset = client.datasets.connect(
    name="capybara",
    source_type="hf",
    huggingface_dataset="trl-lib/Capybara",
)

job = client.training_jobs.create(
    base_model="Qwen/Qwen3-0.6B",
    dataset_id=dataset.id,
    output_name="qwen-0.6b-capybara",
    method="sft_text",
    hyperparameters={
        "learning_rate": 2e-5,
        "num_epochs": 1,
        "max_seq_length": 2048,
    },
    gpu_type="L4-24GB",
    gpu_count=1,
)

job.wait(poll_interval=15)
job.download(output_dir="./checkpoints")

SFT takes no reward function — the labeled completions are the training signal.

Hyperparameters

Parameter	Default	What it does
`learning_rate`	`2e-5`	AdamW step size. Higher than RL; `≈1e-4` is common when training LoRA adapters.
`num_epochs`	`1`	Passes over the dataset. Overridden by `max_steps` if both set.
`max_steps`	`null`	Total gradient steps. Set this to cap a long run or smoke-test.
`max_seq_length`	model default	Tokens per example. Truncate longer sequences. Big OOM lever.
`packing`	`false`	Pack multiple short examples into one sequence for throughput.
`dataset_text_field`	`"text"`	Column holding the text, for plain-text datasets. Ignored for chat data.
`lora_rank`	`null`	Train LoRA adapters of this rank instead of full finetuning. `null` means full finetune.
`lora_alpha`	`null`	LoRA scaling factor. Defaults to `lora_rank` when unset.
`load_in_4bit`	`false`	QLoRA: load the base in 4-bit (NF4) and train LoRA adapters on top. Requires `lora_rank`. Single-GPU.
`use_unsloth`	`false`	Opt-in Unsloth memory optimization. Off by default; single-GPU only.

LoRA and QLoRA

SFT uses the same adapter path as GRPO. Set lora_rank to train low-rank adapters instead of the full model, and add load_in_4bit for QLoRA (a 4-bit NF4 base with adapters on top), which fits larger models on a single GPU.

# QLoRA SFT on a single GPU
hyperparameters={
    "learning_rate": 1e-4,
    "lora_rank": 16,
    "lora_alpha": 32,
    "load_in_4bit": True,
}

See LoRA and QLoRA for the full rules (4-bit requires lora_rank; QLoRA is single-GPU; 16-bit LoRA is multi-GPU capable).

Where to go next

GRPO

Reinforce correctness with RL after an SFT pass.

Datasets

Connect a text or chat dataset.

Submit a job

The full training_jobs.create schema.

Deploy your checkpoint

Serve the trained model with an OpenAI-compatible API.

Get started

Training

Deployments

Evaluations

CLI

Volumes

Dataset format

Submitting a job

Hyperparameters

LoRA and QLoRA

Where to go next

GRPO

Datasets

Submit a job

Deploy your checkpoint

​Dataset format

​Submitting a job

​Hyperparameters

​LoRA and QLoRA

​Where to go next

GRPO

Datasets

Submit a job

Deploy your checkpoint

Dataset format

Submitting a job

Hyperparameters

LoRA and QLoRA

Where to go next