Sırdaş

Dataset for Llama, Mistral and local models

JSONL in TRL and Unsloth's format, ready to fine-tune a model on your own machine.

Sent to servers: 0 BNothing left your device

About this tool

If you are fine-tuning an open model on your machine or on a rented GPU, the documents should never pass through someone else's service just to become a text file. Here they do not.

The conversation format produces the "messages" field with its roles, which TRL, Unsloth and llama-factory load with no translation in between. The plain text format is for continued pretraining on your own material.

How it works

  1. 1

    Drop the PDFs.

  2. 2

    Pick the conversation format.

  3. 3

    Download the JSONL and load it with load_dataset.

Questions

Does it work for LoRA and QLoRA?
Yes: the dataset format is independent of the fine-tuning technique. LoRA changes how weights are updated, not how data is read.
What about DPO or preference data?
We do not generate it, deliberately: a chosen/rejected pair requires producing two answers, which means calling a model with your documents. That would break the one thing we promise.
Can I automate this for hundreds of files?
Yes, with the CLI: npx sirdas dataset *.pdf --split -d ./ds. Same engine.

More tools