A compact, curated DPO (Direct Preference Optimization) dataset of 7k chosen/rejected pairs, carefully selected from multiple high-quality sources. Ideal for running DPO experiments locally — small enough to fine-tune in under an hour on consumer GPUs while producing meaningfully aligned models.
UltraFeedback — The default DPO preference set to run after any SFT pass
Nectar — Reward-model training with 7-way ranked responses
Anthropic HH-RLHF — Safety-focused preference training (helpfulness and harmlessness)
HelpSteer2 — Training reward models with fine-grained quality ratings
Frequently asked questions
Can I use DPO Mix 7K commercially?
Yes — DPO Mix 7K is released under Apache 2.0, a permissive license that allows commercial use, including training models you ship in a product. Check the dataset card for attribution requirements before release.
How much data does DPO Mix 7K contain, and do I need all of it?
DPO Mix 7K contains 7k Pairs. You rarely need all of it: for style and format fine-tuning, a few hundred to a few thousand examples are enough — load a slice (e.g. split="train[:1000]") and scale up only if quality plateaus.
What is DPO Mix 7K best used for?
A small, balanced DPO starter set. It belongs to the Preference (RLHF / DPO) section of our dataset hub, where you'll find alternatives and complementary sets.