Skip to content

Configuration Reference

easylora uses a config-driven design. All settings are defined in a TrainConfig object (Pydantic v2) which can be loaded from YAML or JSON files.

Config File Format

model:
  base_model: "meta-llama/Llama-3.2-1B"
  torch_dtype: "auto"

data:
  dataset_name: "tatsu-lab/alpaca"
  format: "alpaca"
  max_seq_len: 2048

lora:
  r: 16
  alpha: 32
  target_modules: "auto"

training:
  epochs: 3
  batch_size: 4

output:
  output_dir: "./output"

{
  "model": {
    "base_model": "meta-llama/Llama-3.2-1B"
  },
  "data": {
    "dataset_name": "tatsu-lab/alpaca",
    "format": "alpaca"
  }
}

Sections

model -- ModelConfig

Field Type Default Description
base_model str required HF model ID or local path
tokenizer str base_model Tokenizer ID (if different from model)
trust_remote_code bool false Trust remote code in model/tokenizer
torch_dtype str "auto" "auto", "bf16", "fp16", "fp32"
load_in_4bit bool false QLoRA 4-bit quantisation
load_in_8bit bool false 8-bit quantisation
device_map str "auto" Device placement strategy

Quantisation exclusivity

load_in_4bit and load_in_8bit are mutually exclusive. Setting both to true will raise a validation error.

data -- DataConfig

Field Type Default Description
dataset_path str null Local file path (JSON/JSONL/CSV/Parquet)
dataset_name str null HuggingFace Datasets identifier
subset str null Dataset subset/config name
split str "train" Dataset split
text_field str "text" Column name for raw format
format str "raw" "raw", "alpaca", or "chatml"
max_seq_len int 2048 Maximum sequence length (32--131072)
val_split_ratio float 0.0 Fraction for validation split

Dataset source

You must provide either dataset_path (local file) or dataset_name (HF hub). Providing neither raises a validation error.

lora -- LoRAConfig

Field Type Default Description
r int 16 LoRA rank
alpha int 32 LoRA alpha (scaling = alpha/r)
dropout float 0.05 LoRA dropout
target_modules str or list "auto" Modules to adapt
bias str "none" "none", "all", or "lora_only"
task_type str "CAUSAL_LM" PEFT task type
modules_to_save list null Extra modules to train fully

optim -- OptimConfig

Field Type Default Description
lr float 2e-4 Learning rate
weight_decay float 0.01 Weight decay
betas tuple (0.9, 0.999) Adam betas
warmup_ratio float 0.03 Warmup fraction
scheduler str "cosine" LR scheduler type

training -- TrainLoopConfig

Field Type Default Description
epochs int 3 Number of training epochs
batch_size int 4 Per-device batch size
grad_accum int 4 Gradient accumulation steps
logging_steps int 10 Log every N steps
eval_steps int null Evaluate every N steps
save_steps int null Save checkpoint every N steps
gradient_checkpointing bool true Enable gradient checkpointing
max_steps int -1 Max steps (-1 = use epochs)

output -- OutputConfig

Field Type Default Description
output_dir str "./output" Output directory
run_name str null Run name for logging
save_total_limit int 3 Max checkpoints to keep
push_to_hub bool false Push to HuggingFace Hub
hub_repo_id str null Hub repository ID
hub_private bool true Make hub repo private
allow_overwrite bool false Allow overwriting output dir

repro -- ReproConfig

Field Type Default Description
seed int 42 Random seed
deterministic bool false Fully deterministic mode

Loading Configs in Python

from easylora.config import load_config

config = load_config("config.yaml")

CLI Overrides

Use --set to override any config value:

easylora train --config config.yaml --set model.base_model=gpt2 --set optim.lr=1e-5