Skip to content
Dataset By Humans
LLM training data

Conversation dataset format for chatbot fine-tuning (with examples)

The JSONL chat format used to fine-tune chatbots, what makes a good customer service dataset, and how many conversations you really need.

By Dataset By HumansPublished 2 min read

In this article

Short answer: a conversation dataset for chatbot fine-tuning is a JSONL file with one dialogue per line, each holding a messages array of {role, content} turns (system, user, assistant), plus optional labels such as intent and outcome. Quality matters far more than size: 500–2,000 realistic, well-labelled conversations about your own use cases usually beat tens of thousands of generic ones.

The standard format

Almost every fine-tuning stack (OpenAI, Hugging Face TRL, Axolotl, LLaMA-Factory) accepts a chat format close to this:

{"messages": [
  {"role": "system", "content": "You are the support assistant for an online shop. Be brief and polite."},
  {"role": "user", "content": "my order DH-48213 hasn't arrived after 5 days"},
  {"role": "assistant", "content": "Sorry about the wait. Order DH-48213 is at our warehouse and delayed by high volume. It should arrive within 2 days."},
  {"role": "user", "content": "I need it by Saturday"},
  {"role": "assistant", "content": "I've upgraded it to free express delivery, so it will arrive before Friday."}
]}

For training classifiers and for analysis, keep labels next to the messages:

{"id": "csc-000781", "intent": "order-not-received", "sentiment": ["negative", "neutral", "negative", "positive"], "outcome": "resolved", "messages": [...]}

Our datasets ship with both: labels for analytics and a clean messages export for fine-tuning.

What makes a customer service dataset good

  1. Your intents, not generic ones. Order tracking for a marketplace is different from card disputes at a bank.
  2. Hard cases on purpose. Angry customers, fraud attempts, refund edge cases, requests the bot must refuse or escalate.
  3. Natural user language. Typos, missing diacritics, abbreviations and code-switching, the way customers really write.
  4. Consistent agent behaviour. Every assistant turn follows the same policy and tone guide, or the model learns contradictions.
  5. No real personal data. Invented names, numbers and addresses only.
  6. Labels you can trust. Intent, outcome and sentiment checked by a reviewer, not guessed by a model.

Real logs, open datasets or written conversations?

Exported chat logs look like free data, but they carry personal data, reflect your current bot's mistakes, and rarely include the edge cases you need. Open conversational datasets are useful for prototypes; check the license, because many are research-only, and check whether the text was translated or generated by another model. Human-written conversations built from your scenarios cost more per record but need far fewer records, and they are safe to train on.

How many conversations do you need?

  • Prototype: 100–300 conversations covering your top intents
  • Production fine-tune: 500–2,000, balanced across intents with 15–25% hard cases
  • Evaluation set: 100–300 held-out conversations that never go into training

Measure after each round and order more only for the intents where the bot still fails.

Getting a dataset made

We write conversational AI datasets in Vietnamese and English with trained writers and former support agents, from scenarios we design with you. Send your intents and industry and we will reply with sample conversations within 48 hours.

conversational ai datasetscustomer service datasetchatbot training datafine-tuning