Add or import examples
Type them in the form, paste a CSV with prompt and response columns, or drop an existing JSONL in any of the three formats.
Datasets
LLM Fine-tuning Dataset Builder
A supervised fine-tuning set is a list of conversations. This builder gives you a form for system, user and assistant turns (multi-turn supported), imports what you already have in CSV or JSONL, validates every row, estimates tokens, and writes the exact JSONL shape each trainer expects: OpenAI chat format, Alpaca instruction/input/output, or ShareGPT conversations.
Loading the tool…
How it works
Validate
Every example is checked: roles alternate, no empty turns, the last turn is the assistant, token length under the limit you set.
Export
Choose the target format; a split option writes train and validation files with a seed.
Formats, limits and the numbers that matter
Formats
OpenAI: {"messages": [{"role": ..., "content": ...}]}. Alpaca: {"instruction", "input", "output"}. ShareGPT: {"conversations": [{"from": "human"|"gpt", "value": ...}]}. TRL's SFTTrainer reads the first; Axolotl reads all three.
Tokens
Counted with the cl100k tokenizer as a proxy; Llama and Mistral tokenizers run 10-20 percent higher on English and much higher on Hindi or Urdu.
How many examples
For a style or format task, 200-1,000 clean examples move a 7B model noticeably; for knowledge, fine-tuning is the wrong tool and retrieval is the right one.
Common questions
Should the system prompt be in every example?
Include it if inference will use it; the model learns the pairing. Keep it identical across examples unless the task is prompt-following.
Can I mask the prompt from the loss?
That is a trainer setting (completion-only loss in TRL); the data format is the same.
Is my data stored?
Only in this browser's storage so you can continue later; nothing is sent anywhere.