Quickstart

From zero to a fine-tuned model in 10 minutes. This guide walks you through the dashboard, step by step.

1. Create your account

Go to parsyn.progatis.com and click Sign Up. Fill in your name, email, and a password (8 characters minimum), then confirm. You'll be redirected to the login page.

Sign in with your credentials. You land on the Dashboard, which shows your resource counts (datasets, models, jobs, workers) and quick action buttons.

2. Get compute ready

You need GPU power to train models. Two options:

Option A: Use Parsyn Workers (no setup needed)

Go to the Credits page from the sidebar. You'll see your current balance at the top. Switch to the Buy tab to see available credit packs with their prices. Pick a pack and click Purchase to complete the payment through Stripe.

Credits are used to pay for GPU time on Parsyn's managed fleet. Pricing varies by GPU type. Check the Pricing tab to see rates per GPU. That's all the setup you need. When you create a training job later, just select "Parsyn Workers" and the platform handles the rest.

Option B: Connect your own GPU

If you have a machine with an NVIDIA GPU (CUDA 11.8+, 8 GB+ VRAM), you can connect it to the platform. Training on your own hardware is free.

Go to the Workers page and click Add Worker. Enter a name (e.g., "my-rtx-4090"), an optional description, and select the worker type:

• Both: Training and inference (recommended)
• Fine-Tuner: Training only
• Prompter: Inference only

Click Create. The platform generates an API key and displays it in a dialog. Copy this key immediately. It's shown only once.

On your GPU machine, install the worker client and start it:

pip install parsyn-worker

export PLATFORM_URL="wss://parsyn.progatis.com/ws/worker"
export WORKER_ENROLLMENT_KEY="enroll_xyz789..."
export GPU_DEVICE="cuda:0"

parsyn-worker start

After a few seconds, the worker appears in your Workers list with a green status dot. It runs a GPU benchmark automatically, then sits idle waiting for tasks.

3. Upload a dataset

Prepare a JSONL file with prompt and completion fields:

{"prompt": "What is gradient descent?", "completion": "Gradient descent is an optimization algorithm that iteratively adjusts model parameters by moving in the direction of steepest decrease of the loss function."}
{"prompt": "Explain backpropagation.", "completion": "Backpropagation computes the gradient of the loss with respect to each weight by applying the chain rule, propagating error signals from the output layer back through the network."}
{"prompt": "What is overfitting?", "completion": "Overfitting occurs when a model learns the training data too well, capturing noise rather than the underlying pattern, which degrades performance on unseen data."}

Go to the Datasets page and click Upload Dataset. In the dialog:

1. Enter a name (e.g., "ml-fundamentals")
2. Add an optional description
3. Click the file input to select your .jsonl file (or drag and drop it)
4. Click Upload

The file uploads and the dataset appears in your list. Click on it to open the detail view, where you can preview rows, check the schema, and compute statistics.

4. Register a base model

Go to the Models page and click Add Model. In the dialog:

1. Enter a name (e.g., "gpt2-base")
2. Enter the HuggingFace model ID: gpt2
3. Add an optional description
4. Click Create

The model appears in your list with a "Base Model" badge. The worker will download it from HuggingFace when training starts.

5. Launch a training job

Go to the Training page and click New Training Job. The creation form has several sections:

Basic configuration

1. Enter a job name (e.g., "gpt2-ml-fundamentals")
2. Select your dataset from the dropdown
3. Select your base model
4. Set the core parameters: learning rate, batch size, number of epochs, and max sequence length. The defaults are reasonable for a first run.

Advanced options (optional)

Expand the advanced section if you want to configure LoRA, mixed precision, gradient accumulation, learning rate scheduler, evaluation, or checkpointing. See Training Configuration for details on each parameter.

Worker selection

Choose how the platform assigns compute:

• Auto: The platform picks the best available worker based on GPU benchmark scores. You can toggle whether to include your workers, Parsyn Workers, or both.
• My Workers: Select specific workers from your connected machines. No credits consumed.
• Parsyn Workers: Use managed GPUs. Credits are deducted based on GPU type and training duration.
• Mixed: Select from both your workers and Parsyn's fleet.

Click Create. The job appears in the training list with a "pending" status. Click the play button or open the job detail and click Start Job.

The platform assigns the job to an available worker. Once running, the detail page shows live metrics: loss curve, learning rate, GPU utilization, throughput, and ETA. These update in real time through WebSocket.

6. Test the fine-tuned model

Once training completes (green "completed" badge), the fine-tuned model is saved to storage. Go to the Chat page.

Select your fine-tuned model from the dropdown at the top. Type a message in the input field and press Send. The model generates a response, streamed token by token.

You can adjust generation settings (temperature, top_p, max tokens) in the settings panel on the right. Enable Compare Mode to see responses from the fine-tuned model and its base model side by side.

What's next

• Dataset Management for larger files, format conversion, and preprocessing tools
• Training Configuration for LoRA, mixed precision, and hyperparameter recipes
• Distributed Training to scale across multiple GPUs with federated learning
• Worker Setup for production-grade worker deployment on your infrastructure