Create a fine-tuning dataset from your PDFs
Turn your documents into the JSONL trainers expect, without uploading them to any service.
About this tool
Fine-tuning a model on your own documents starts with a boring task: turning PDFs into a JSONL file with the exact field names the trainer expects. Get one name wrong and the load fails twenty minutes later, with an error that explains nothing.
Here you pick the format and it comes out with the right fields: "text" for continued pretraining, "messages" for conversation, "prompt" and "completion" for instruction. These are TRL's names, which is also what Unsloth and OpenAI fine-tuning eat.
And because the file never leaves your machine, you can do it with medical records, contracts or case files without asking anyone's permission.
How it works
- 1
Drop all the PDFs at once.
- 2
Pick the format and leave anonymization and duplicate cleanup ticked.
- 3
Build and download the .jsonl.
Questions
- How many documents do I need?
- It depends on the trainer, and it is worth reading its docs first: OpenAI fine-tuning asks for at least 10 examples, and Claude 3 Haiku on Bedrock for 32. The minimum lets the run start, which is not the same as getting a good result.
- Does fine-tuning teach the model new facts?
- It mostly teaches format, tone and how to answer. For the model to know specific facts from your documents and cite them, RAG usually works better: retrieve the passage and hand it over at the time. Our PDF to Markdown tool with chunking prepares exactly that.
- What about personal data?
- Anonymization is on by default. A model repeats what it sees: names, phone numbers and emails have been extracted from the text a model was trained on, even from passages that appeared only once.