Skip to main content

Fine-Tune Domain Models Quickly to Unlock Custom AI Productivity

📌 Scenario Overview​

Enterprises and research institutions often possess structured or unstructured domain-specific data, such as medical records, legal regulations, or financial reports. These datasets carry strong industry characteristics, where general-purpose LLMs may underperform. Leveraging CSGHub's model and data hosting capabilities, users can swiftly prepare training data and fine-tune open-source models to create tailored domain-specific LLMs.

  • Target Users: Startups / Industry research institutes
  • Goal: Train a business-aligned language model using industry corpora for tasks like Q&A systems, knowledge extraction, and document generation.

🧭 Step-by-Step Guide​

1. Create a Training Dataset​

  • Log in to your CSGHub deployment and create a new training dataset under your personal or team account. alt text
  • For collaborative management, create an organization and invite members. alt text

2. Upload Domain-Specific Data​

  • Upload training corpora in .jsonl, .txt, or other common formats for supervised fine-tuning (SFT) or other tasks. Add descriptions and tags for easier management and reuse.
  • Multiple data upload methods are supported.

3. Select a Base Model​

  • Browse and select a suitable open-source base model (e.g., DeepSeek, Qwen, Baichuan, InternLM) from the model list. Check its details and license.
  • Use the Model Tree feature on the model details page to review its lineage and derivatives for informed selection. alt text

4. Launch a Fine-Tuning Task​

  1. Create an instance with its name, compute resources and framework.
  2. Wait until ready, then open the framework UI, choose the dataset and training parameters, and start training.
  3. Verify logs and checkpoint output. A running instance does not mean a training job has started.

5. Publish New Model​

  • After finetuning, upload the model to your personal or organizational repository via CLI or Web UI. Release it as a new version with tags and documentation for traceability.

🌟 Key Outcomes​

  • Obtain a fine-tuned domain model with significantly improved comprehension of industry terminology and contexts.
  • Deploy the model in private environments for applications like Q&A bots and document generation.
  • Streamlined fine-tuning workflow accessible even to non-technical teams via Web UI.
  • Clear version control with rollback/upgrade support for continuous optimization.