Zum Inhalt

Quick Start: Run Your First Training

DeepExtension makes it easy to launch your first model training job through a fully visual interface — no coding required.

Depending on your system’s AI platform, we provide preinstalled demo training methods to help you get started.

Available Demo Training Methods

For CUDA（Linux or Windows via WSL) Users

  • GRPO-Demo: A ready-to-use GRPO (Guided Reinforcement with Prompt Optimization) method for logic-aligned fine-tuning.
  • SFT-Demo: A supervised fine-tuning example suitable for small-scale tasks.

For macOS Users (Apple Silicon with MLX)

  • MLX-Demo: A demonstration training method using Apple’s MLX framework, optimized for M-series chips.

Before You Start

As outlined in Process Dependency, make sure the following components are ready:

  • A base model has been added
  • A dataset is available
  • A demo training method (already preinstalled)

We recommend starting with one of the following example datasets:

You’re also welcome to use your own dataset — just make sure it follows the same structure as one of these examples, including matching field names.

Suggested Parameters for Demo Training

Below is a minimal parameter configuration that works reliably for all demo training methods:

Parameter Name Description Recommended Value
Base Model Foundation model for fine-tuning Qwen2.5-1.5B
Dataset Dataset to be used demo-dataset
LORA_RANK LoRA adapter rank 8
LOAD_IN_4BIT Enable 4-bit quantization true
MAX_SEQ_LENGTH Max sequence length 1024
MAX_INPUT_LENGTH Max input length 1024
MAX_CONTENT_LENGTH Max content length 1024
MAX_SAMPLES Max number of training samples 1000
NUM_GENERATIONS Generations per batch 2
MAX_GRAD_NORM Gradient clipping norm 0.5
EPOCHS* Number of training epochs 1
MAX_STEPS Max training steps 10
BATCH_SIZE Training batch size 2
GRAD_ACCUM_STEPS Gradient accumulation steps 2
LEARNING_RATE Initial learning rate 2e-5
WARMUP_STEPS Warmup steps 2
WARMUP_RATIO* Warmup ratio (alternative to warmup steps) 0.01
OUTPUT_DIR Output directory System auto-generated
PromptInputColumn** Input prompt field name
PromptOutputColumn** Output/response field name

* These parameters will be available in the very near future.
** These parameters are not required for all Demo Training Methods.

You can fine-tune these parameters later as you gain more experience.

Monitor Progress

  • Go to Model Training in the main navigation.
  • Launch the job and track its progress.
  • Once the job reaches the Completed state, the output will appear under Customized Models.

Next Step

From here, you can:

DeepExtension — Fast start, full control.