How to train a ChatGPT Model

Before you start

Training ChatGPT requires a Eskra Pro account. Once you upgrade to Pro, you'll have access to our training tools in your dashboard. Training typically takes 10 minutes to a few hours depending on the size of your dataset.

Two ways to train your model

Eskra offers two methods for creating training data. Choose the one that works best for you:

• Flow Map (Recommended for beginners): Answer a business-focused questionnaire and we'll generate training data for you
• Upload Custom Data (Recommended for best results): Create your own training data using ChatGPT for maximum flexibility

Method 1: Using Flow Map

Flow Map is perfect if you want to get started quickly without technical knowledge. Here's how it works:

• Step 1: Navigate to Custom GPT in your Eskra settings
• Step 2: Answer the questionnaire about your business, including your industry, typical customer inquiries, and preferred tone
• Step 3: Flow Map generates training data examples based on your answers
• Step 4: Review and edit the generated examples to ensure they match your style
• Step 5: Click "Start Training" and we'll create and fine-tune your custom ChatGPT model

Method 2: Upload Custom Training Data

For the best results, we recommend creating training data from your actual email conversations. This method gives you complete control and produces highly personalized models.

Step 1: Generate training data with ChatGPT

We provide a special prompt that helps ChatGPT create properly formatted training data for you. Here's what to do:

• Go to chat.openai.com and start a new conversation
• Copy and paste our training prompt (see below)
• Answer ChatGPT's questions about your business and email style
• Provide 10+ examples of actual customer emails and your responses
• ChatGPT will generate a JSONL file with your training data

The training prompt

Copy this entire prompt and paste it into ChatGPT:

You are an AI assistant tasked with helping the user create training data for an email-response model. Begin by asking the user questions to understand their email context, audience, tone, and style. Collect details such as: The user's business or role and who they email, the typical purpose of their emails (customer support, sales, etc.), their usual tone and vocabulary (formal, friendly, technical, etc.), any specific words/phrases to use or avoid, and examples of actual email queries and how they reply. Important: The dataset must contain at least 10 complete examples (an incoming email + the user's reply) to meet minimum requirements for fine-tuning. However, aim to gather 40–50 examples or more, since larger datasets produce much better fine-tuned models. Continue prompting the user for additional examples until they indicate they are finished. Ask one question at a time, wait for the user's response, and adapt follow-up questions based on their answers. Once you have gathered enough information, generate training examples in JSONL format.

Step 2: Upload your training data

Once ChatGPT generates your training data:

• Copy the JSONL output from ChatGPT
• Return to Eskra dashboard and navigate to Training
• Select "Upload Custom Training Data"
• Paste your JSONL data or upload the file
• Our system automatically verifies the format is correct
• Click "Start Fine-Tuning" to begin training your model

Training data requirements

To create a successful fine-tuned model, your training data must meet these requirements:

• Minimum 10 examples: This is the bare minimum required by OpenAI
• Recommended 40-50+ examples: More examples produce significantly better results
• JSONL format: Each line must be valid JSON with a messages array
• Real examples work best: Use actual customer emails and your real responses

After training completes

Once your model is trained, you can:

• Select it as your active model in AI Settings
• Test it by generating replies to your emails
• Compare it against the base ChatGPT model
• Retrain anytime with additional examples to improve results