Fine-Tuning Dataset Builder (Spreadsheet to Chat JSONL)

About the Fine-Tuning Dataset Builder (Spreadsheet to Chat JSONL)

Turn a spreadsheet of examples into the file a fine-tuning job needs. Open a CSV, TSV, Excel or OpenDocument sheet (or paste cells copied from one), say which columns hold the system prompt, the user’s message and the ideal answer — or group rows into multi-turn conversations by an id column — and download chat-format JSONL, the layout OpenAI’s fine-tuning and Hugging Face TRL read, or prompt–completion, Gemini tuning, ShareGPT and Alpaca files for other trainers.

Before you pay for a training run, the checks list what would waste it: empty answers, turns in the wrong order, copies, the same question answered two ways, examples longer than the token limit you set (counted with the LLM Token Counter’s tokenizers), unbalanced labels and too few examples. Split off a validation file — stratified by a label if you like, with a seed so the split repeats — mask emails, phone numbers and other personal data first, and see the token totals with a cost estimate from your price. You can also open an existing JSONL or JSON dataset to check it or convert it to another format. Everything runs in your browser; nothing is uploaded.

Common uses

• Turning a sheet of questions and approved answers from a support team into a first fine-tuning set.
• Converting chat logs exported one message per row into multi-turn training examples.
• Checking a dataset for empty answers, copies and unbalanced labels before paying for a training run.
• Converting a dataset between chat JSONL, prompt–completion, Gemini, ShareGPT and Alpaca layouts for another trainer.
• Masking customer emails and phone numbers before the data goes to a training service.

The file formats

• Chat messages: {"messages": [{"role": "system" | "user" | "assistant", "content": "…"}]}, one example per line. It is OpenAI’s fine-tuning format (supervised fine-tuning) and Hugging Face TRL’s conversational format (dataset formats); Axolotl’s chat_template type (conversation datasets) and Unsloth (datasets guide) read it too. Train only on the last answer adds OpenAI’s "weight": 0 to the earlier answers of a conversation (fine-tuning best practices).
• Prompt and completion: {"prompt": […], "completion": [{"role": "assistant", "content": "…"}]}. TRL’s SFTTrainer then computes the loss on the completion only, by default (SFT trainer).
• Gemini tuning: {"systemInstruction": {"parts": [{"text": "…"}]}, "contents": [{"role": "user" | "model", "parts": [{"text": "…"}]}]}, for supervised tuning of Gemini models on Google Cloud (prepare the data).
• ShareGPT: {"conversations": [{"from": "system" | "human" | "gpt", "value": "…"}]}, which Axolotl maps with field_messages and Unsloth converts with standardize_sharegpt.
• Alpaca: {"instruction", "input", "output"}, plus "system", like Stanford Alpaca’s alpaca_data.json, which is one JSON array (Stanford Alpaca). It holds one turn; multi-turn conversations are left out unless you write the earlier turns as "history", which only some trainers read.

What the checks look for

• Errors (left out of the files by default): no user message or no answer, a system message after the conversation has started, a conversation that starts with the assistant, rows with an unknown role or no conversation id.
• Warnings: two user messages or two answers in a row; a conversation that ends with the user (models learn only from answers); exact copies; examples over the token limit, which trainers cut at the end — OpenAI: “Examples longer than the default are truncated … which removes tokens from the end of the training example”; labels where the largest has at least 3 times as many examples as the smallest; one answer given in 30% or more of the examples (at least 5). OpenAI’s guide warns that if 60% of the answers say “I cannot answer this” when only 5% should, “you will likely get an overabundance of refusals”.
• Tips: the same question with different answers (a model can only be as consistent as its data), and the size of the set. OpenAI’s minimum is 10 examples, and it reports improvements from 50–100; Google suggests starting with 100 for Gemini (about Gemini tuning), and Unsloth at least 100 rows.

Length limits to choose from: 65,536 tokens per example for OpenAI’s fine-tunable GPT-4.1 and GPT-4o models, 131,072 for Gemini tuning, and 1,024, the default max_length of TRL’s SFTTrainer.

Tokens and the cost estimate

Tokens per example = the tokens of its messages, counted with the tokenizer you choose, + 4 per message + 3 for the roles and separators — the way OpenAI’s cookbook counts chat messages (counting tokens). Chat templates of open models add a similar few tokens. Claude and Gemini tokenizers are not published, so those counts are labelled estimates.

Billed training tokens = the tokens of the training file × the epochs (Google: “the number of tokens in your training dataset by the number of epochs”); an example over the limit counts only up to the limit, because the rest is cut off. Auto epochs follow OpenAI’s data-preparation cookbook: 3, raised for small sets until examples × epochs reaches 100 (at most 25), and lowered to 25,000 ÷ examples for very large sets (data preparation and analysis). The cost is billed tokens × your price per million ÷ 1,000,000.

Limitations

• Text conversations only: tool calls, tool results, images and audio are not supported, and imported records that contain them are left out (the checks list their line numbers).
• Token counts use the tokenizer you choose plus a fixed allowance for roles; the exact count depends on the model’s chat template, so leave a margin under a hard limit.
• The checks find mechanical problems. They cannot tell whether an answer is right or written the way you want your model to answer: read a sample of the examples yourself.
• Masking finds emails, phone numbers, card and bank numbers, identity numbers, IP addresses and the words you list; it cannot find every name or street address in free text.
• Up to 200,000 rows are read from a sheet; formulas are read as the values the file last saved.

Privacy

Your sheet, the examples and the files you download are built in your browser and never uploaded. Counting tokens downloads only the tokenizer you choose, once.