I am trying to create a bot to answer FAQs about my product. I have tried to train my data on pretrained models like ada, curie, and davinci. My dataset has only 30 questions and I suspect this might be the reason of the poor response from the bot. Could you please suggest me the best way to create a Q&A bot using chatgpt api? The bot should be trained on my private data. I have loads of information in a word file in paragraph format. I want the bot to learn these and answer questions. Please help.
2 Likes
Fine tuning is not the correct way to teach the model new data, for that you can either include the information in the prompt or you can try embeddings, see Embeddings
For a Q&A bot using the ChatGPT API, here’s a simple approach:
- Prepare Data:
- Convert information in your Word file to a question-answer format.
- Break down paragraphs into concise questions and corresponding answers.
- Increase Dataset Size:
- Expand your dataset beyond 30 questions for more diverse training.
- Train the Model:
- Utilize the ChatGPT API to fine-tune on your dataset.
- Use prompts like “Given this question, what’s the best answer?”
- Iterative Testing:
- Test the bot with various questions to refine its responses.
- If needed, continue to add more data and retrain iteratively.
By expanding your dataset and fine-tuning on specific Q&A formats, you can enhance the bot’s performance in answering questions related to your product.
Use embeddings and a vector database to store your vectors that represent your data.