Create a Text Generation Experiment
Create an LLM fine-tuning experiment and start training. The creation page has two pages — Configuration (page 1) → Publish (page 2) — navigated with the Next button at the bottom.
Step 1. Configuration (Page 1)
Click Create experiment, then configure the experiment basics and base model.
| Field | Required | Description |
|---|---|---|
| Experiment name | ✓ | Name identifying the experiment |
| Local queue selection | ✓ | Local queue to run the experiment |
| Description | — | Experiment description |
| Select template | — | Load a predefined training-settings template |
| Total GPUs | ✓ | Number of training GPUs (slider 1–8) |
| Base model | ✓ | Base model to fine-tune (search · All/Base/Fine-tuning tabs) |
Click Next to move to the Publish page (Cancel discards).
Step 2. Publish — Dataset & Training Model (Page 2)
| Field | Required | Description |
|---|---|---|
| Dataset | ✓ | Select a dataset from Thaki Datasets / Custom Datasets tabs and Add |
| Method | ✓ | Training method (SFT, Full Fine-Tuning, etc.) |
| Training model | ✓ | Model used for training |
| Teacher | — | Teacher model for distillation (card selection) |
Step 3. Publish — Hyperparameters
Hyperparameters settings has Basic / Resource / Advanced / Early stopping / Expert sections. Key Basic settings:
| Field | Description |
|---|---|
| Max Steps | Total training steps |
| Epochs | Number of passes over the dataset |
| Learning Rate | Weight update step size (default 5e-5) |
| Per Device Batch Size | Samples processed per GPU per step |
| Gradient Accumulation Steps | Steps to accumulate before updating |
| Max Sequence Length | Maximum sequence length of training samples |
Step 4. Publish — Resource / Advanced Settings, then Create
| Section | Description |
|---|---|
| Resource settings | CPU · Memory · Distributed Strategy (DDP, etc.) · Number of Nodes |
| Advanced settings | Warmup Steps · LR Scheduler · Logging Steps · Precision |
| Early stopping settings | Early-stopping conditions |
| Expert settings | Additional parameters for advanced users |
Review your settings and click Create to create and start the experiment.