llm-datasets

Why we included this project

If you are getting ready to fine-tune or align your own model, this is the reference you want open in a tab. It is a living directory of SFT and RL/preference datasets grouped by domain, and each entry notes sample counts, whether the data includes reasoning traces, and the license. The same repository pulls together post-training tools and guides, so you can go from picking a starting mixture for a chat assistant or a reasoning model to actually assembling the pipeline. The opening section on dataset quality, covering accuracy, diversity, and complexity, gives a quick sanity check before you commit to any collection. It is a curated resource rather than deployable software, but it will save you real time when you are building your own training data.