Configs - Monte Platform

A training config is a YAML file that states how a run trains: the base model, the method, LoRA, and the optimizer settings. The run states its training data, step count, and starting point with flags, so one config serves many runs.

Write a Training Config

triage-sft.yaml trains Qwen/Qwen3-8B with SFT and a LoRA adapter:

model: Qwen/Qwen3-8B
method: sft

lora:
  rank: 64
  alpha: 128

batch_size: 32
learning_rate: 1.0e-4
max_seq_length: 4096
checkpoint_every: 50

Parameter	Description
`model`	The base model, as a Hugging Face ID from Models
`method`	`sft`, `dpo`, `grpo`, or `distillation`
`lora`	A LoRA adapter. Leave the block out for a full fine-tune. An empty block takes the defaults below.
`lora.rank`	The adapter rank (default 16)
`lora.alpha`	The scale of the adapter update (default twice the rank)
`lora.target_modules`	The layers that get an adapter (default all linear layers)
`lora.dropout`	The dropout on the adapter input (default 0.0)
`batch_size`	Rows, pairs, or prompts in one step
`learning_rate`	The peak learning rate
`max_seq_length`	The longest sequence in tokens
`checkpoint_every`	Save a checkpoint every this many steps (default 50)

The adapter’s rank must not exceed the --max-adapter-rank of the base deployment that serves it (default 64). Each method adds its own keys under its name, such as grpo.group_size. The method pages list them: SFT, DPO, GRPO, and Distillation.

Use a Config in a Run

The file name is the config name, so configs/triage-sft.yaml is --config triage-sft. --config also takes a path to a YAML file.

monte train support-triage --config triage-sft --data support-replies \
  --steps 200 --from-base

--model replaces model for one run. The run records the hash of the file, so an edit to the file makes a new version of the config.