Importing Models and Data#

This guide covers how to import existing models and data as W&B artifacts using the nemotron CLI. This is useful when you want to:

  • Use a pre-existing checkpoint from another training run

  • Import data prepared outside of the standard pipeline

  • Connect external assets to the W&B artifact lineage system

Prerequisites#

  • W&B configuration in env.toml (see Execution through NeMo-Run):

    [wandb]
project = "nemotron"
entity = "YOUR-TEAM"

  • Or provide --project and --entity CLI flags

Model Import#

Import model checkpoints as W&B artifacts for use in downstream training stages.

Commands#

# Import pretrain model checkpoint
uv run nemotron nano3 model import pretrain /path/to/model_dir --step 10000

# Import SFT model checkpoint
uv run nemotron nano3 model import sft /path/to/model_dir --step 5000

# Import RL model checkpoint
uv run nemotron nano3 model import rl /path/to/model_dir --step 2000

Options#

Option

Description

--step, -s

Training step number (optional)

--name, -n

Custom artifact name (default: nano3/<stage>/model)

--project, -p

W&B project (overrides env.toml)

--entity, -e

W&B entity (overrides env.toml)

Examples#

# Import with custom artifact name
uv run nemotron nano3 model import pretrain /lustre/checkpoints/model --step 50000 --name my-pretrain-model

# Import to different W&B project
uv run nemotron nano3 model import sft /path/to/sft_checkpoint --project other-project --entity my-team

Data Import#

Import data directories as W&B artifacts for use in training stages.

Commands#

# Import pretrain data (expects blend.json file)
uv run nemotron nano3 data import pretrain /path/to/blend.json

# Import SFT data (expects directory with blend.json)
uv run nemotron nano3 data import sft /path/to/sft_data_dir

# Import RL data (expects directory with manifest.json)
uv run nemotron nano3 data import rl /path/to/rl_data_dir

Expected Directory Structures#

Pretrain: Direct path to blend.json file

/path/to/blend.json

SFT: Directory containing blend.json

/path/to/sft_data_dir/
├── blend.json
├── train.npy
├── valid.npy
└── ...

RL: Directory containing manifest.json

/path/to/rl_data_dir/
├── manifest.json
├── train.jsonl
├── val.jsonl
└── test.jsonl

Options#

Option

Description

--name, -n

Custom artifact name (default: nano3/<stage>/data)

--project, -p

W&B project (overrides env.toml)

--entity, -e

W&B entity (overrides env.toml)

Examples#

# Import SFT data with custom name
uv run nemotron nano3 data import sft /lustre/data/sft_v2 --name my-sft-data

# Import RL data to different project
uv run nemotron nano3 data import rl /path/to/rl_data --project alignment-project

Model Evaluation#

uv run nemotron nano3 model eval

Note: Model evaluation is coming soon.

Using Imported Artifacts#

After importing, artifacts can be referenced in training commands via --art.<slot> (see CLI Framework):

# Use imported model in SFT training
uv run nemotron nano3 sft --art.model my-pretrain-model:latest --run YOUR-CLUSTER

# Use imported data in training
uv run nemotron nano3 pretrain --art.data my-pretrain-data:v1 --run YOUR-CLUSTER

CLI Reference#

Model Commands#

uv run nemotron nano3 model --help
uv run nemotron nano3 model eval --help
uv run nemotron nano3 model import --help
uv run nemotron nano3 model import pretrain --help
uv run nemotron nano3 model import sft --help
uv run nemotron nano3 model import rl --help

Data Import Commands#

uv run nemotron nano3 data import --help
uv run nemotron nano3 data import pretrain --help
uv run nemotron nano3 data import sft --help
uv run nemotron nano3 data import rl --help

Further Reading#