Config file

This page explains every field in the training section of the YAML config.

The SDK reads a single YAML file that contains both data_generation and training. run_training(config_path=...) reads the training section from that file.

File format

Your YAML file is a run config that includes top-level keys and a nested training section:

run_steps: ["data_generation", "training"]
pipeline_provider: "local"
metric_name: "custom"

training:
  metric: {}
  dataset: {}
  prompt_template: ""
  training: {}
  output: {}
  model: {}

Configuration structure

The training section has these parts:

metric: output type and class list (if multi-class)
dataset: where to load the dataset created by data generation
prompt_template: the exact format the model is trained to follow
training: hyperparameters and performance options
output: where to write artifacts + optional pushes (Hub/Luna Studio/object store)
model: base model + LoRA settings + model download options

`metric`

Field	Type	Required	Default (base config)	Notes
`metric.type`	`string`	Yes	`""`	One of `boolean`, `multi_class`.
`metric.classes`	`list[string] \| null`	Conditionally	`[]`	Required when `type` is `multi_class`. Order defines the class index mapping.

`dataset`

Field	Type	Required	Default (base config)	Notes
`dataset.name`	`string`	Yes	`""`	For Hub datasets: `org/repo`. For local: typically the generated dataset folder name.
`dataset.split`	`string`	Yes	`"train"`	Training split.
`dataset.test_split_name`	`string`	Yes	`"test"`	Holdout split.
`dataset.train_label_column`	`string`	Yes	`"label"`	Label column for training.
`dataset.test_label_column`	`string`	Yes	`"label"`	Label column for evaluation.
`dataset.local`	`bool`	Yes	`false`	When true, load from disk using `dataset.name` as a path-like identifier.
`dataset.local_path`	`string`	Yes	`""`	Base path where local datasets exist/extract to.
`dataset.pull_from_object_store`	`bool`	No	`false`	Download dataset from object store before training.
`dataset.object_store_bucket`	`string \| null`	If object store	`"${LUNA_OBJECT_STORE_BUCKET:-luna-ft}"`	Required when `pull_from_object_store` is true.
`dataset.object_store_blob`	`string \| null`	If object store	`"generated_data"`	Required when `pull_from_object_store` is true.

Point training at the generated dataset

run_data_generation(...) returns a dataset path/name. After generation completes, set:

training.dataset.name to that value, then run training.

`prompt_template`

prompt_template is required.

It must reference your dataset columns using {variable} placeholders.
For boolean metrics, the model should respond with "true" or "false".
For multi-class metrics, the model should respond with a single class-key token ("0", "1", …).

Tip: your template variables are validated against dataset columns during preflight.

`training` (hyperparameters)

Field	Type	Required	Default (base config)
`num_train_epochs`	`int`	Yes	`5`
`per_device_train_batch_size`	`int`	Yes	`1`
`gradient_accumulation_steps`	`int`	Yes	`8`
`learning_rate_multiplier`	`float`	Yes	`1.0`
`max_seq_length`	`int`	Yes	`4096`
`warmup_steps`	`int`	Yes	`30`
`weight_decay`	`float`	Yes	`0.01`
`max_grad_norm`	`float`	Yes	`1.0`
`logging_steps`	`int`	Yes	`40`
`torch_compile`	`bool`	Yes	`false`
`fp16`	`bool`	Yes	`false`
`bf16`	`bool`	Yes	`true`
`optim`	`string`	Yes	`"adamw_8bit"`
`lr_scheduler_type`	`string`	Yes	`"linear"`
`seed`	`int`	Yes	`3407`
`use_wandb`	`bool`	Yes	`true`

`output`

Field	Type	Required	Default (base config)	Notes
`output.model_name`	`string`	Yes	`""`	Artifact folder name and (optionally) Hub model name.
`output.local_path`	`string`	Yes	`"data/"`	Root directory for artifacts.
`output.push_to_hub`	`bool`	Yes	`true`	If true, pushes model artifacts to Hugging Face. Requires `HF_TOKEN` when enabled.
`output.push_to_object_store`	`bool`	No	`false`	Upload artifacts tarball to the configured object store.
`output.object_store_bucket`	`string \| null`	If object store	`"${LUNA_OBJECT_STORE_BUCKET:-luna-ft}"`	Required when `push_to_object_store` is true.
`output.object_store_blob`	`string \| null`	If object store	`"finetuned_models"`	Required when `push_to_object_store` is true.

Environment variables for pushes

Hugging Face: HF_TOKEN (required when push_to_hub: true)
W&B: WANDB_API_KEY (optional; warning if missing when use_wandb: true)

`model`

Field	Type	Required	Default (base config)	Notes
`model.name`	`string`	Yes	`"meta-llama/Llama-3.2-3B-Instruct"`	Base model repo id.
`model.peft_model_map`	`map[string,string]`	No	`{}`	Map base-model → LoRA weights repo id.
`model.use_unsloth_tokenizer`	`bool`	Yes	`true`	Use Unsloth Tokenizer when available.
`model.lora_r`	`int`	Yes	`16`	LoRA rank.
`model.lora_alpha`	`int`	Yes	`16`	LoRA alpha.
`model.pull_model_from_object_store`	`bool`	No	`false`	Download base model + adapters from object store.
`model.model_object_store_bucket`	`string \| null`	If object store	`"${LUNA_OBJECT_STORE_BUCKET:-luna-ft}"`	Required when `pull_model_from_object_store` is true.
`model.model_object_store_blob_base`	`string \| null`	If object store	`"base_models"`	Base model location.
`model.model_object_store_blob_adapters`	`string \| null`	If object store	`"adapters"`	Adapters location.
`model.model_local_path`	`string \| null`	No	`""`	Local directory for models (also used for local-only runs).

Overview

Get Started

Observability

Evaluation Metrics

Signals

AI Assistant

Luna Studio

Experiments

Agent Control

Annotations

Integrations

Security

References

File format

Configuration structure

`metric`

`dataset`

Point training at the generated dataset

`prompt_template`

`training` (hyperparameters)

`output`

Environment variables for pushes

`model`

​File format

​Configuration structure

​metric

​dataset

​Point training at the generated dataset

​prompt_template

​training (hyperparameters)

​output

​Environment variables for pushes

​model

File format

Configuration structure

`metric`

`dataset`

Point training at the generated dataset

`prompt_template`

`training` (hyperparameters)

`output`

Environment variables for pushes

`model`