Configuration

All training parameters live in config.yaml → runtime_info.input.training. The block generates the LLaMA-Factory YAML from these values at runtime. This page documents the fields that matter most and the gotchas behind them.

Core fields

Field	Meaning
`stage`	Training stage (`sft`)
`finetuning_type`	`full` for full fine-tuning
`deepspeed`	DeepSpeed config path (`ds_z3_config.json` = ZeRO-3)
`template`	Chat template (see below)
`cutoff_len`	Max sequence length (tokens)
`rope_scaling`	RoPE scaling method (see below)
`max_samples`	Upper bound applied by LLaMA-Factory after loading
`output_dir`	Run name / checkpoint directory
`per_device_train_batch_size`	Batch size per GPU
`gradient_accumulation_steps`	Gradient accumulation steps
`learning_rate`	Learning rate
`num_train_epochs`	Number of epochs
`lr_scheduler_type`	LR schedule (e.g. `cosine`)
`warmup_ratio`	Warmup fraction of total steps
`save_strategy`	Checkpoint schedule: `steps`, `epoch`, or `no`
`save_steps`	Checkpoint interval when `save_strategy: steps`
`save_only_model`	Save only model weights, without resumable optimizer state

The effective global batch size is per_device_train_batch_size × gradient_accumulation_steps × n_gpus.

Template

Match the template family to the model: Qwen3 uses qwen3 / qwen3_nothink; Qwen3.5 uses qwen3_5 / qwen3_5_nothink. Select a thinking-capable variant only when the dataset intentionally preserves reasoning content. The active Qwen3.5 profile uses qwen3_5. See Scaffolds.

Long context

Enable a model-supported RoPE scaling method such as yarn when cutoff_len exceeds the base model's native context window (max_position_embeddings or the documented context length). There is no universal 32K threshold: compare the requested cutoff with the selected model's own limit.

The checked-in Qwen3.5 profile uses rope_scaling: null: its native context already covers the configured 128K cutoff, and this architecture ignores the older top-level RoPE scaling option.

Checkpointing

save_only_model: true skips saving optimizer state — it saves disk but prevents resuming.
save_strategy: steps uses save_steps; epoch saves at epoch boundaries.
resume_from_checkpoint: null — set it to a checkpoint directory path to resume.
A non-empty output_dir is rejected by default. Destructive replacement requires both overwrite_output_dir: true and the explicit runtime acknowledgement SFT_ALLOW_OVERWRITE_OUTPUT=1.
A relative output_dir writes to artifacts/model/<basename>; an absolute path is honored as-is by both training and the dashboard.

Performance toggles

Field	Effect
`bf16: true`	bf16 mixed-precision training
`enable_liger_kernel: true`	Liger fused kernels
`use_unsloth_gc: true`	Unsloth gradient checkpointing
`flash_attn: fa2`	FlashAttention-2, built against the installed PyTorch

Experiment tracking

experiment.wandb_mode controls reporting: disabled generates report_to: none; offline logs locally without an API key; online requires WANDB_API_KEY in the private runtime environment.

Tunable parameters

The hyperparameters worth tuning between runs (all under runtime_info.input.training / .conversion):

Parameter	Range / meaning
`learning_rate`	1e-5 to 1e-3
`num_train_epochs`	Number of epochs
`per_device_train_batch_size`	Batch size per GPU
`gradient_accumulation_steps`	Gradient accumulation steps
`max_instances`	Harbor conversion cap or ready-made-source `num_samples`

Common failure modes

Symptom	Fix
Single-GPU run on an 8-GPU node	Ensure `FORCE_TORCHRUN=1` (set by `train.sh`) and `n_gpus_per_node` is correct
Dataset key points to an old LF filename	Let `train.sh` update the mapping, or pick a new `conversion.data_name`
Requested `cutoff_len` exceeds the model's native context window	Enable a RoPE scaling method supported by that model, such as `yarn`
Non-empty output directory rejected	Rename `output_dir`; only intentional destructive replacement should set both overwrite controls
WandB online mode without an API key	Export `WANDB_API_KEY`, or switch to `offline`/`disabled`

Configuration

On this page