CLI Reference

CLI Reference#

This page contains a complete reference of all CLI commands available in Oumi.

For detailed guides and examples of specific areas (training, inference, evaluation, etc.), please refer to the corresponding user guides in the documentation.

CLI Overrides#

Any Oumi command which takes a config path as an argument (train, evaluate, infer, etc.) can override parameters from the command line. For example:

oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml \
  --training.max_steps 20 \
  --training.learning_rate 1e-4 \
  --data.train.datasets[0].shuffle true \
  --training.output_dir output/smollm-135m-sft

Oumi uses OmegaConf to parse the configs from YAML files, and to parse the command line overrides. OmegaConf allows a Pythonic specification of parameters to override with dot-separated syntax, as seen above. Note that for lists (ex. data.train.datasets), you can specify the index either with brackets ([0]) or dot notation (.0).

With OmegaConf, you can set the value of an entire dictionary or list, in addition to overriding individual primitive values. For example:

# Override one entry in the list. Note that the new dict is merged with the existing
# one, so the existing value of `"dataset_name": "yahma/alpaca-cleaned"` is kept.
oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml \
  --data.train.datasets[0] '{"shuffle": True, "sample_count": 100}'

# Override the list, in this case to add a new entry.
# Note that we redundantly specify an existing entry in the list here.
oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml \
  --data.train.datasets '[{"dataset_name": "yahma/alpaca-cleaned"}, {"dataset_name": "CohereForAI/aya_dataset"}]'

Warning

OmegaConf doesn’t readily support adding/deleting entries in a list from command line overrides using index notation. Instead, you need to set the value of the entire list, or modify the YAML config.

Warning

By default, when you override a dict value with another dict, the items from both will be merged, preferring the value from the overriding dict in case of an existing key. This is equivalent to Python dict merging behavior, ex. new_dict = {**old_dict, **override_dict}

Training#

For a detailed guide on training, see Training.

Evaluation#

For a detailed guide on evaluation, see Evaluation.

Inference#

For a detailed guide on inference, see Inference.

Judge#

For a detailed guide on judging, see LLM Judge.

Launch#

For a detailed guide on launching jobs, see Running Jobs on Clusters.

Distributed#

For a detailed guide on distributed training, see Training.

Data Synthesis#

For a detailed guide on data synthesis, see Data Synthesis.

Dataset Analysis#

For a detailed guide on dataset analysis, see Dataset Analysis.

Hyperparameter Tuning#

For a detailed guide on hyperparameter tuning, see Hyperparameter Tuning.

Quantization#

For a detailed guide on model quantization, see Quantization.

Environment#

This command is a great tool for debugging!

oumi env will list relevant details of your environment setup, including python version, package versions, and Oumi environment variables.