#Fine-Tuning a Customer Support Chatbot: Data Preparation Questions and Examples

7 messages · Page 1 of 1 (latest)

craggy crow
#

Hey everyone,
I am currently working on fine-tuning a model for customer support. However, I am running into some issues and am in need of some guidance. I am using the suggestions from the case study on the customer support chatbot, which can be found here: https://beta.openai.com/docs/guides/fine-tuning/case-study-customer-support-chatbot
One question I have is about how to prepare my data. In the descriptions, it states that there should be a single input (prompt) and output (completion), but in the case study, there are a couple of questions and answers in one row. For example:
\nCustomer: <message1>
\nAgent: <response1>
\nCustomer: <message2>
\nAgent: <response2>

Do I need to include "Customer:" and "Agent:" in every row of my data, or is it enough to just add the text? And how does the model know where the question and the answer are? Should I add one input and one output for a row, or I can use multiple?

Can anyone clarify which way is the right way to prepare my data? And could you provide a couple of examples of how to format the data?

Thank you!

OpenAI API

An API for accessing new AI models developed by OpenAI

ornate harbor
#

Yes you should include it to keep it on track. It depends on what you're trying to tune it towards. You wouldn't want to cut off half of a meaningful conversation. Stick to the case study. Talk to the bot and figure out what you'd like to be tuned

inland quail
#

I'd suggest playing around with a trial of Microsoft virtual agent, to better get your head around how a conversation needs to be structured.

craggy crow
craggy crow
ornate harbor
#

@craggy crow Think about how you want to structure your prompt. It makes sense to think that 1 chat = 2 messages but there's two issues with that

  1. Sometimes there's context needed from something said 2-3 conversation pairs back
  2. As a result, you would need to resend the whole conversation to maintain a "memory". Which in turns makes your prompts dynamic
ornate harbor
#

So considering those 2 points I'd ask myself how I can deliver quality prompts