one-click fine-tuning of DeepSeek-R1 distill models - Platform For AI

Supported models

Model Gallery supports LoRA SFT for the six distill models. The following table lists minimum configurations with default parameters:

Distill model	Base model	Training method	Minimum configuration
DeepSeek-R1-Distill-Qwen-1.5B	Qwen2.5-Math-1.5B	LoRA supervised fine-tuning	1 x A10 (24 GB video memory)
DeepSeek-R1-Distill-Qwen-7B	Qwen2.5-Math-7B		1 x A10 (24 GB video memory)
DeepSeek-R1-Distill-Llama-8B	Llama-3.1-8B		1 x A10 (24 GB video memory)
DeepSeek-R1-Distill-Qwen-14B	Qwen2.5-14B		1 x GU8IS (48 GB video memory)
DeepSeek-R1-Distill-Qwen-32B	Qwen2.5-32B		2 x GU8IS (48 GB video memory)
DeepSeek-R1-Distill-Llama-70B	Llama-3.3-70B-Instruct		8 x GU100 (80 GB video memory)

Train the model

Go to the Model Gallery page.
1. Log on to the PAI console.
2. In the upper-left corner, select a region.
3. In the left pane, click Workspaces. On the Workspaces page, click a workspace name.
4. In the left pane, choose QuickStart > Model Gallery.
On the Model Gallery page, click the DeepSeek-R1-Distill-Qwen-7B model card to go to the details page.

This page displays deployment, training details, SFT data format, and invocation methods.

Click Train in the upper-right corner and configure the following key parameters:

Dataset Configuration: Upload prepared data to an OSS bucket.
Computing Resources: Minimum configurations are listed in Supported models. Adjusting hyperparameters may require more memory.

Hyperparameters: Adjust these LoRA SFT hyperparameters based on your data and resources. For details, see Fine-tune LLMs.

Hyperparameter	Type	Default value (for 7B model as an example)	Description
learning_rate	float	5e-6	Controls weight adjustment magnitude.
num_train_epochs	int	6	Number of training epochs (dataset iterations).
per_device_train_batch_size	int	2	Samples per GPU per iteration. Higher values increase efficiency and memory usage.
gradient_accumulation_steps	int	2	The number of gradient accumulation steps.
max_length	int	1024	Max tokens per sample.
lora_rank	int	8	LoRA dimension.
lora_alpha	int	32	LoRA scaling factor.
lora_dropout	float	0	LoRA dropout rate. Randomly drops neurons during training to prevent overfitting.
lorap_lr_ratio	float	16	LoRA+ learning rate ratio (λ = ηB/ηA). Uses different rates for adapter matrices A and B. Set to 0 for standard LoRA.

Click Train. The training page shows job status and logs.
- On success, the model is registered in AI Asset Management - Models for deployment. See Register and manage models.
- On failure, click next to Status or check the Task log tab. For common errors, see FAQ and Model Gallery FAQ.
- Metric Curve shows the loss progression.
After training, click Deploy to create an EAS service. Invocation follows the original distill model. See the model detail page or Deploy DeepSeek-V3 and DeepSeek-R1 models.

Billing

Model Gallery training uses DLC, billed by job duration. Resources stop automatically when jobs end. See Billing of DLC.

FAQ

Why does my Model Gallery training job fail?

Cause: max_length too small. Data exceeding this limit is discarded:

Solution: Increase max_length. If too much data is discarded, training or validation datasets may become empty, causing failure:
Error: failed to compose dlc job specs, resource limiting triggered, you are trying to use more GPU resources than the threshold

Solution: Training is limited to 2 simultaneous GPUs. Wait for ongoing jobs to finish, or submit a ticket to increase quota.
Error: the specified vswitch vsw-**** cannot create the required resource ecs.gn7i-c32g1.8xlarge, zone not match

Solution: The requested instance type is unavailable in the current zone. Try one of these:
- Leave vSwitch empty. DLC auto-selects one based on inventory.
- Switch to a different instance type.

How do I download the trained model from Model Gallery?

Set the model output path to an OSS directory when creating the training job, then download the model from OSS.

How can I improve poor model performance after fine-tuning?

Try the following approaches:

Use a larger model with better baseline performance, such as DeepSeek or Qwen3 series with higher parameter counts.
Refine your prompts.
Increase max_tokens.
Break complex tasks into smaller subtasks for the model to handle separately.

What is the required OSS path structure for custom model uploads?

When uploading a fine-tuned model to OSS for deployment, the model directory must contain all required configuration files. Missing files like generation_config.json can cause deployment failures.

Required OSS path structure:

oss://your-bucket/model-path/
├── config.json                    # Model architecture configuration
├── generation_config.json         # Generation parameters (required)
├── tokenizer_config.json          # Tokenizer configuration
├── tokenizer.json                 # Tokenizer vocabulary
├── special_tokens_map.json        # Special tokens mapping
├── pytorch_model.bin              # Model weights (or .safetensors)
├── adapter_config.json            # LoRA adapter config (for LoRA models)
└── adapter_model.bin              # LoRA adapter weights (for LoRA models)

Example generation_config.json:

{
  "bos_token_id": 151643,
  "do_sample": true,
  "eos_token_id": [
    151645,
    151643
  ],
  "pad_token_id": 151643,
  "repetition_penalty": 1.0,
  "temperature": 0.3,
  "top_k": 20,
  "top_p": 0.8,
  "transformers_version": "4.37.0"
}

Solutions if files are missing:

Copy from base model: If generation_config.json is missing, copy it from the original base model directory used for fine-tuning.
Create manually: Create a generation_config.json file with default parameters (see example above). Adjust bos_token_id, eos_token_id, and pad_token_id to match your tokenizer.
Upload complete model directory: When setting the model output path during training (see How do I download the trained model from Model Gallery?), ensure the training job saves all configuration files, not just model weights.

Note

For LoRA fine-tuned models, keep both base model files and adapter files in the same directory, or configure separate paths for base model and adapter in your deployment configuration.

Platform For AI:Fine-tune DeepSeek-R1 distill models