sft/automodel#

This step runs supervised fine-tuning (SFT) on a Hugging Face (HF) format model from a JSON Lines (JSONL) chat dataset, using the NeMo AutoModel library. The same step supports full fine-tuning and low-rank adaptation (LoRA) tuning, controlled by the peft parameter.

Syntax#

nemotron steps run sft/automodel \
    [-c <config-name-or-path>] \
    [-r <run-profile> | -b <batch-profile>] \
    [-d] \
    [--force-squash] \
    [<dotlist-overrides>...] \
    [<passthrough-args>...]

Refer to the Nemotron Steps CLI Reference for the shared flag set.

Configuration Files#

The step ships two configuration files under src/nemotron/steps/sft/automodel/config/.

File

Purpose

default.yaml

Full-shape training on Qwen/Qwen3-30B-A3B with the AutoModel TrainFinetuneRecipeForNextTokenPrediction recipe.

tiny.yaml

Short validation run with five training steps and a 64-record dataset slice from HuggingFaceH4/ultrachat_200k.

Pass either name with -c:

$ nemotron steps run sft/automodel -c tiny
$ nemotron steps run sft/automodel -c default

Inputs and Outputs#

Direction

Artifact Type

Required

Description

Consumes

training_jsonl

Yes

Instruction data in JSONL with a messages field in OpenAI chat format.

Produces

checkpoint_hf

—

Hugging Face checkpoint directory. The output is a full model checkpoint when peft=null, and a LoRA adapter directory when peft=lora.

Supported Models#

The manifest declares three reference base models. Other Hugging Face causal language models that load through AutoModelForCausalLM.from_pretrained also work.

Model

Minimum GPUs

Default

Notes

Qwen/Qwen3-30B-A3B

8

Yes

Mixture-of-experts (MoE) base model used by config/default.yaml.

meta-llama/Llama-3.1-8B-Instruct

4

No

Common dense baseline for single-node SFT and LoRA.

mistralai/Mistral-7B-Instruct-v0.3

2

No

Use this model when GPU count is small or iteration speed matters.

Override the base model from the command line:

$ nemotron steps run sft/automodel -c default \
    model.pretrained_model_name_or_path=meta-llama/Llama-3.1-8B-Instruct

Step Parameters#

The manifest declares one Nemotron-specific parameter. Pass it as a dotlist override after the options.

peft=VALUE#

Selects adapter-style training instead of full fine-tuning. Set this value to lora to train a LoRA adapter, or to null for full SFT.

Choices: lora, null.

Default: null.

Example: peft=lora

You can also override any key from the compiled YAML configuration. The frequently used keys include the following.

step_scheduler.max_steps=N#

The maximum number of optimizer steps for the run.

Example: step_scheduler.max_steps=200

step_scheduler.global_batch_size=N#

The global batch size across all data-parallel workers.

Example: step_scheduler.global_batch_size=64

dataset.path_or_dataset_id=<id-or-path>#

The Hugging Face dataset identifier or a local path that resolves to a JSONL chat dataset.

Example: dataset.path_or_dataset_id=/data/my-instructions.jsonl

checkpoint.checkpoint_dir=PATH#

The directory where the AutoModel recipe writes checkpoints.

Example: checkpoint.checkpoint_dir=/output/qwen-sft

Strategies#

The manifest records four operator strategies for sft/automodel.

  • When the run has one or two GPUs, or memory is tight, set peft=lora and start from a Mistral-class model.

  • When the run has three or four GPUs and the chosen model fits comfortably, consider full fine-tuning only when the resulting checkpoint size and iteration speed remain acceptable.

  • When the dataset already uses OpenAI chat-format JSONL, skip data_prep/sft_packing and train directly from training_jsonl.

  • When you need an immediately deployable HF checkpoint, keep the safetensors save format and the consolidated HF output layout that config/default.yaml sets.

Common Errors#

chat_template_missing#

Cause: the tokenizer for the chosen model does not include a chat template, so the AutoModel recipe cannot render the messages field into prompt-and-completion form.

Recovery: choose a tokenizer with chat-template support, or convert the data to prompt-and-completion format before training.

oom#

Cause: GPU memory is exhausted during forward or backward passes.

Recovery: set peft=lora, reduce step_scheduler.global_batch_size, or move to a smaller model.

Command Examples#

Run the tiny validation configuration locally:

$ nemotron steps run sft/automodel -c tiny

Compile the default configuration on a Lepton profile without submitting the job:

$ nemotron steps run sft/automodel -c default -r lepton_sft_automodel --dry-run

Submit an attached LoRA run on a Lepton profile with a smaller base model:

$ nemotron steps run sft/automodel -c default -r lepton_sft_automodel \
    peft=lora \
    model.pretrained_model_name_or_path=mistralai/Mistral-7B-Instruct-v0.3 \
    step_scheduler.max_steps=500

Submit a detached full SFT on Slurm and write checkpoints to a shared scratch directory:

$ nemotron steps run sft/automodel -c default -b slurm_sft_automodel \
    peft=null \
    step_scheduler.global_batch_size=64 \
    checkpoint.checkpoint_dir=/lustre/runs/qwen-sft