Config reference

Purpose

Reference for all TrainConfig fields and defaults. Any YAML field not set explicitly falls back to the default listed here.

When to use

• While creating a new training config.
• When checking which settings are safe on limited hardware.
• When validating reasoning/truncation/loss settings.

Config load behavior

1. Token resolution order:
   - `--hf_token` CLI flag
   - `HF_TOKEN` environment variable
   - YAML `hf_token` field
2. Default resolution:
   - A default is applied only when the field is omitted in YAML.

Schema (YAML shape)

run_name: "run"
model_name: "unsloth/Qwen3-4B-bnb-4bit"
dataset_id: "your-hf-username/your-dataset"
max_length: 2048

Options by section

Identity

Field	Default	Description
run_name	"run"	Subdirectory name under output_dir/ and adapter_base_dir/
output_dir	"outputs"	Root dir for trainer checkpoints and logs
adapter_base_dir	"adapters"	Root dir for saved LoRA adapter
hf_token	null	HF access token (prefer HF_TOKEN env var)

Model

Field	Default	Description
model_name	"unsloth/Qwen3-4B-bnb-4bit"	HF repo id or local path
chat_template	"qwen3"	Unsloth template key (qwen3 works for Qwen3 and Qwen3.5)

Valid chat_template keys commonly used:

Key	Use for
"qwen3"	Qwen3 and Qwen3.5
"qwen3-thinking"	Qwen3 with explicit thinking template
"qwen3-instruct"	Qwen3 instruct without thinking
"qwen-2.5"	Qwen2.5 family

Dataset

Field	Default	Description
dataset_id	""	HF dataset id or local path (required)
dataset_split	"train"	Split passed to load_dataset
messages_field	"messages"	Column with conversation list
dataset_schema	"auto"	auto / messages / prompt_response

Reasoning / thinking

Field	Default	Description
reasoning_field	"reasoning_content"	Canonical field for <think> content
reasoning_keys	null	Alternative keys normalized into reasoning_field
extract_think_tags	true	Extract inline <think>...</think> from assistant content
think_mode	"keep"	keep preserve reasoning, drop remove reasoning
think_max_tokens	0	Cap tokens per reasoning block (0 = no cap)
think_role	"think"	Role name for separate-think-message datasets
think_loss	"all"	all / answer_only / answer_plus_think

think_loss behavior:

Value	Gradient scope
all	Full assistant span (<think> + answer)
answer_only	Tokens after </think>
answer_plus_think	Think content + answer (excluding literal tags)

Sequence / truncation

Field	Default	Description
max_seq_length	2048	Positional allocation; set >= max_length
max_length	2048	Max tokens/sample; primary VRAM lever
truncate_side	"left"	Fallback token-level truncation side

Precision / hardware

Field	Default	Description
load_in_4bit	true	BNB NF4 quantization
attn_implementation	"sdpa"	sdpa or flash_attention_2
fp16	true	Float16 path (commonly Qwen3)
bf16	false	BFloat16 path (commonly Qwen3.5)

LoRA

Field	Default	Description
lora_r	16	LoRA rank
lora_alpha	32	Scaling factor (commonly 2 * lora_r)
lora_dropout	0.0	LoRA dropout
lora_target_modules	null	null -> all 7 projection layers
gradient_checkpointing	"unsloth"	unsloth / true / false

Training

Field	Default	Description
per_device_train_batch_size	1	Keep at 1 for limited VRAM setups
gradient_accumulation_steps	8	Effective batch = batch size x accumulation
learning_rate	2e-4	Peak LR for AdamW
warmup_ratio	0.05	LR warmup fraction
max_steps	1000	Total optimizer steps
logging_steps	20	Log interval
save_steps	200	Checkpoint interval
seed	3407	Random seed
optim	"adamw_8bit"	adamw_8bit or adamw_torch

Loss masking

Field	Default	Description
assistant_roles	null	Roles that carry loss (null -> ["assistant"])
drop_if_no_assistant	true	Drop samples with no assistant turn

Validation rules

1. Required dataset source:
   - `dataset_id` must be non-empty.
2. Precision flags:
   - `fp16` and `bf16` should not be enabled at the same time.
3. Sequence budget:
   - high `max_length` increases VRAM pressure and can trigger OOM on 8 GB GPUs.
4. Positional allocation:
   - `max_seq_length` should be >= `max_length`.

Edge cases / limitations

WARNING

dataset_id is required. Empty dataset_id cannot be resolved into a train dataset.

• fp16 and bf16 should not be enabled simultaneously.
• Aggressive max_length on 8 GB VRAM can cause OOM.

Related

• Quickstart
• Dataset pipeline
• Reasoning control