Why fine-tune on 200K samples when 6K will do?

What it does

Deita provides automatic data-selection pipelines, pre-filtered datasets, and resulting chat models for LLM alignment. It ships trained scorer models that rate instruction-following examples by complexity and quality, then uses them to carve massive pools into lightweight subsets—as small as 6K samples—for supervised fine-tuning and DPO. The released chat models are built on Mistral-7B and LLaMA-13B bases.

The interesting bit

The project bets that dataset curation matters more than dataset volume. Its Mistral-based DEITA-7B-v1.0 was trained on just 6K SFT plus 10K DPO examples, yet the README tables place it within striking distance of Zephyr-7B-β and Starling-7B—rivals that consumed hundreds of thousands of examples. That is a lot of GPU hours left on the table.

Key highlights

• Ships ready-made 6K and 10K SFT datasets, plus a 300K “SOTA pool” for custom selection.
• Includes trained complexity and quality scorers to rank your own instruction data.
• Released chat models span Mistral-7B and LLaMA-13B variants.
• Pipelines are configurable and support VLLM for faster scoring inference.
• Licensing is split: datasets are MIT, models are Apache 2.0 or LLaMA 2, and scorer weights are under the LLaMA license.

Verdict

Worth a look if you are building alignment pipelines and suspect most of your instruction dataset is noise. Look elsewhere if you need a general pre-training or RL framework—this is strictly about curating and tuning instruction data.