Fine-tune with the API or CLI

Summary

Supplements knowledge (Injects domain knowledge)

Learns to perform tasks (Follows instructions)

Performs tasks better (Aligns with human preferences)

Input data

10 million+ tokens

Unlabeled domain text

Over 1,000 entries

High-quality "question-answer" pairs

100+ sets

"Better-worse" response pairs for the same instruction

Core objective

Domain adaptation. Learns specialized vocabulary and facts.

Teaches the model conversation formats and task execution capabilities.

Makes model outputs better align with human values and preferences.

Learning method

Self-supervised learning (Predicts the next word)

Supervised learning (Imitates the ground truth)

Direct preference learning (Increases the probability of good responses and decreases the probability of bad responses)

Model stage

Typically before SFT

After CPT and before DPO

Typically after SFT, as the final step for alignment.

Scenarios

• The model needs to acquire new capabilities

• Achieving optimal global performance.

• Optimizing model performance for specific scenarios.

• For cost-sensitive and time-sensitive scenarios.

Training time

Longer, with slower convergence.

Shorter, with faster convergence.

n_epochs

(Number of epochs) [Required]

Data size < 10,000: 3–5

Data size > 10,000: 1–2

Integer

How many times the model iterates over the full training set.

More epochs increase training time and cost.

learning_rate

(Learning rate)

Use the recommended default value provided by Model Studio.

Float

Controls the step-size for weight updates during training.

• Too high: training may diverge.

• Too low: slow convergence with minimal improvement.

freeze_vit

(Freeze visual backbone)

Adjust as needed

Boolean

Freezes the visual backbone so its weights are not updated during training. Applies only to Qwen-VL models.

batch_size

(Batch size) [Required]

Use the recommended default value provided by Model Studio.

Integer

Number of training examples per iteration. Small values significantly increase training time. Defaults vary by model.

eval_steps

(Evaluation steps)

Adjust as needed

Integer

Step interval for evaluating training accuracy and loss.

Controls how often Validation Loss and Validation Token Accuracy are reported.

logging_steps

(Logging steps)

Adjust as needed

Integer

Step interval for logging training progress.

lr_scheduler_type

(Learning rate scheduler)

Recommended linear/Inverse_sqrt

String

Strategy for adjusting the learning rate during training.

Each strategy is described in Fine-tune a model in the console.

max_length

(Sequence length) [Required]

8192

Integer

Maximum sequence length (in tokens) per training example. Longer examples are discarded.

How to convert between tokens and characters.

max_split_val_dataset_sample

(Max validation set samples)

Use the recommended default value provided by Model Studio.

Integer

When "validation_file_ids" is not set, the validation set that is automatically split by Model Studio contains a maximum of 1,000 entries.

This parameter has no effect when "validation_file_ids" is set.

split

(Training set ratio)

Use the recommended default value provided by Model Studio.

Float

If "validation_file_ids" is not set, Model Studio splits the training file: 80% for training, 20% for validation.

When "validation_file_ids" is set, this parameter has no effect.

warmup_ratio

(Warm-up ratio)

Use the recommended default value provided by Model Studio.

Float

Fraction of training steps used for learning rate warm-up (linear ramp from near-zero to the target rate).

Stabilizes early training by limiting large weight updates.

Too high: resembles a low learning rate with minimal performance change.

Too low: resembles a high learning rate and may degrade performance.

This parameter has no effect if the learning rate scheduler is set to constant.

weight_decay

(Weight decay)

Use the recommended default value provided by Model Studio.

Float

L2 regularization strength. Helps preserve generalization, but excessively high values reduce fine-tuning effectiveness.

Parameters for efficient fine-tuning (supportsefficient_sft and dpo_lora)

lora_rank

(LoRA rank)

64

Integer

Rank of the LoRA low-rank matrices. Higher ranks can improve results but slightly increase training time.

lora_alpha

(LoRA alpha)

Use the recommended default value provided by Model Studio.

Integer

Scaling factor for combining base model weights with the LoRA correction.

Larger alpha: more weight to task-specific LoRA updates.

Smaller alpha: more retention of base model knowledge.

lora_dropout

(LoRA dropout)

Use the recommended default value provided by Model Studio.

Float

Dropout rate for LoRA low-rank matrices. The default balances generalization; overly large values diminish fine-tuning effectiveness.

Parameters for publishing model parameter snapshots (forefficient_sft and sftonly)

save_strategy

(Snapshot save strategy)

It can be set to epoch or steps.

• When set to steps, you can adjust the save interval by setting the save_steps parameter.

String

Strategy for saving model parameter snapshots (checkpoints): epoch (after each epoch) or steps (at a specified step interval).

save_steps

(Save steps)

If you need to modify it manually, set it to an integer multiple of the eval_steps parameter.

Integer

Interval, in training steps, between snapshot saves.

save_total_limit

(Snapshot save limit)

10

Integer

Maximum snapshots to retain. Older snapshots are auto-deleted when the limit is reached.