Configuration File

PMetal supports YAML configuration files as an alternative to CLI flags. Generate a sample config with pmetal init.

Example

model:
  name: "meta-llama/Llama-3.2-1B"
  dtype: bfloat16
  device: gpu
  max_seq_len: 2048

lora:
  r: 16
  alpha: 32.0

training:
  learning_rate: 2e-4
  batch_size: 1
  epochs: 3
  lr_schedule: cosine
  warmup_steps: 100
  gradient_accumulation_steps: 4
  max_grad_norm: 1.0
  weight_decay: 0.01

dataset:
  path: "./train.jsonl"
  format: jsonl

Usage

pmetal train --config training.yaml

CLI flags override config file values when both are specified.

Sections

model

Field	Type	Default	Description
name	string	required	HuggingFace ID or local path
dtype	string	bfloat16	Data type: float32, float16, bfloat16
device	string	gpu	Device: gpu, cpu
max_seq_len	integer	0	Max sequence length (0 = auto)

lora

Field	Type	Default	Description
r	integer	16	LoRA rank
alpha	float	32.0	LoRA scaling factor

training

Field	Type	Default	Description
learning_rate	float	2e-4	Learning rate
batch_size	integer	1	Micro-batch size
epochs	integer	1	Training epochs
lr_schedule	string	cosine	LR schedule type
warmup_steps	integer	100	Warmup steps
gradient_accumulation_steps	integer	4	Gradient accumulation
max_grad_norm	float	1.0	Gradient clipping
weight_decay	float	0.01	AdamW weight decay

dataset

Field	Type	Default	Description
path	string	required	Dataset file path
format	string	auto	Format: jsonl, json, parquet, csv

See Also

• pmetal init — Generate a sample config
• pmetal train — CLI training parameters