Chat Test

Load or connect a trained model and check the fine-tuning results by chatting with it directly.

Loading a Model into the Inference Server

Load a trained model directly to test it.

• Model path: Enter the path of the model to test (e.g., /@team/.AppData/fine_tuning/runs/model)
• Adapter path (optional): If you're using only the adapter weights trained separately, such as via LoRA, specify the path here
• Template: Automatically detected by default

After entering the path, click the Start button to launch the inference server. The Connect setting then switches to Inference server.

Connecting to an Ollama Server

You can also test directly with a model already registered in Ollama.

1. In Connect, select Server Ollama or Local Ollama (my PC).
2. Load the model list with the refresh button (↺).
3. Select the model to test (e.g., gemma4:12b).

If you select Custom API, enter the API address and model name yourself. Click Reset to clear the conversation.

Checking Results Through Conversation

Enter a message in the input box at the bottom and send it with Ctrl+Enter (or the send button); the user message and the AI's response are displayed as a chat. Ask questions relevant to your training goal to check the quality and format of the responses.

Good to Know

• A model that has only completed PT (pre-training) may produce continuation-style responses rather than instruction-following conversation. To check instruction-following quality, test with a model that has also completed SFT.
• Start is unavailable while the model path is empty. Specify the model path even when you use an adapter.

Related Documents

• Evaluation
• Export