Hủy
Cloud Insights

All documents

Model Fine-tuning

FPT Cloud Server
FPT Container
FPT Network
FPT Storage
FPT Cloud Backup & DR
FPT Database Engine
FPT Integration
FPT DevSecOps Services
FPT Monitoring
FPT Security
AI Infrastructure
AI Inference
FPT AI Inference
AI Studio
AI Marketplace
Billing
    Create Pipeline
    Create Pipeline
    Updated on 25 Apr 2025
    Print: Export: PDF

    Data sample : Access GitHub (link) to get the sample data for use

    Access the Model Fine-tuning service and choose Pipeline Management tab, click button "Create pipeline"

    file

    1. Select base model

    file

    • We currently offer base models for fine-tuning, including:
    Base Model Description
    Llama-3.1-8B Base language model, 8B parameters, versatile, ideal for fine-tuning
    Llama-3.2-1B Lightweight base model, 1B parameters, fast, efficient, suitable for edge use
    Llama-3.2-8B-Instruct Instruction-tuned model, 8B parameters, optimized for dialogue and tasks
    Llama-3.2-11B-Vision-Instruct Multimodal instruction-tuned model, 11B parameters, optimized for vision-language tasks
    Llama-3.3-70B-Instruct Instruction-tuned LLaMA model, 70B parameters, excels at complex tasks
    Meta-Llama-3-8B-Instruct Instruction-tuned LLaMA model, 8B parameters, optimized for conversational tasks
    Qwen2-0.5B-Instruct Small instruction-tuned model, 0.5B parameters, lightweight and task-efficient
    Qwen2-VL-7B-Instruct Multimodal instruction-tuned model, 7B parameters, efficient vision-language understanding
    Qwen2-VL-72B Multimodal base model, 72B parameters, handles both vision and language
    Qwen2-VL-72B-Instruct Multimodal instruction-tuned model, 72B parameters, vision-language understanding and generation
    Qwen2.5-0.5B-Instruct Updated instruction-tuned model, 0.5B parameters, improved efficiency and task handling
    Qwen2.5-14B-Instruct Instruction-tuned language model, 14B parameters, balanced power and efficiency
    Qwen2.5-32B-Instruct Instruction-tuned language model, 32B parameters, strong at understanding tasks
    Qwen2.5-VL-72B-Instruct Multimodal instruction-tuned model, 72B parameters, excels at vision-language tasks
    Mixtral-8x7B-v0.1 Sparse Mixture-of-Experts model, 8 experts, high efficiency, strong performance
    Mixtral-8x22B-v0.1 Large Mixture-of-Experts model, 8×22B experts, scalable, efficient, powerful reasoning
    Mixtral-8x22B-Instruct-v0.1 Instruction-tuned MoE model, 8×22B experts, excels at following tasks
    DeepSeek-R1 Foundation language model by DeepSeek, versatile, powerful, and open-source
    DeepSeek-R1-Distill-Llama-70B Efficient language model, distilled from LLaMA 70B, optimized performance
    DeepSeek-R1-V3-0324 Advanced multilingual model, latest DeepSeek version, optimized for diverse tasks

    Note: If you want to upload your models, please contact us!

    2. Select data format

    • Supported data formats for fine-tuning currently include:
    Data Format Description Data Structure File Format
    Alpaca Instruction-following format with input, output pairs for supervised fine-tuning tasks {instruction, input, output} json, zip
    Corpus Large structured text collection, used for training and evaluating models {text} json, zip
    ShareGPT Trained on ShareGPT dataset for improved conversational responses multi-turn chats / {conversations [from, value]} json, zip
    ShareGPT_Image Optimized for multimodal (text & image) processing multi-turn chats / {conversations [from, value]} + image_path zip: train.json and folder images
    • Note: Each dataset must contain at least 50 samples.

    More details about Data format, please visit here: https://fptcloud.com/en/documents/model-fine-tuning/?doc=select-data-format

    3. Prepare the dataset

    • You have two ways to upload the Training/Evaluation dataset:

      • Upload a file: Supported formats: ZIP (zip 2 times), JSON (Maximum: 100MB)

      • Choose a connection: Choose a connection and enter the path to the object within the bucket

        Before selecting a connection, you need to access the Data Hub and create a connection by selecting a data source, entering the endpoint URL to the bucket, providing the access key and secret key. You can also refer to the connection creation guide here: https://fptcloud.com/en/documents/data-hub/?doc=create-connection

    4. Training configuration

    file

    • Select a Trainer
    Trainer Description Supported Data Format
    Pre-training Initial training phase using large unlabeled data for language understanding Corpus
    SFT Supervised fine-tuning trainer, aligns model behavior using labeled data Alpaca/ ShareGPT/ ShareGPT_Image
    DPO Direct Preference Optimization trainer, aligns model with human preference signals directly ShareGPT/ ShareGPT_Image

    5. Set Hyperparameters

    Parameters Description Type Supported values Default value
    learning_rate Learning rate for training float [0.00001-0.001] 0.00001
    batch_size Batch size for training. In case of distributed training, this is batch size on each device int updating 1
    epochs Number of training epochs int updating 1
    gradient_accumulation_steps Number of updates steps to accumulate the gradients for, before performing a backward/update pass int updating 4
    checkpoint_steps Number of training steps before two checkpoint saves if save_strategy="steps". int updating 1000
    max_sequence_length Max input length, longer sequences will be cut-off to this value. int updating 2048
    finetuning_type Which parameter mode to use. enum[string] lora/full lora
    distributed_backend Backend to use for distributed training. Default is ddp enum[string] ddp/deepseed ddp
    deepspeed_zero_stage Stage to apply DeepSpeed ZeRO algorithm. Only apply when distributed_backend=deepspeed enum[int] 1/2/3 1

    More details about Hyperparameters, please visit here: https://fptcloud.com/en/documents/model-fine-tuning/?doc=set-hyperparameters

    6. Select Training node

    • Supports both single-node and multi-node configurations, with a maximum of 16 nodes

    7. Trigger Fine-tuning

    Trigger Description
    Manual User-initiated fine-tuning.
    Scheduled Automated fine-tuning based on a set schedule.

    8. Additional Configurations

    file

    • Upload Test Data (ZIP, JSON max 100MB).
    • Set Number of checkpoint
    • Enable Email Notifications: Notifications will be sent via email to the account that creates the pipeline

    9. Review & Finalize

    file

    1. Enter Pipeline Name
      • Default format: ft_[base model]_[timestamp]
      • Editable with a 50-character limit
    2. Enter Pipeline Description (Max 200 characters)
    3. Review Configurations before finalizing.
    4. Save as Template (Optional)
      • Default format: ft_[base model]_[timestamp]_template
      • Editable within a 100-character limit
    5. Click Finish to create the pipeline.

    You can manually start the pipeline by clicking the Start button or set it to run on a scheduled basis.

    © 2025 FPT Cloud. All Rights Reserved.