UltraFeedback
Large-scale, fine-grained preference dataset. It contains 64k prompts with multiple model responses rated by GPT-4. Essential for RLHF and DPO.
from datasets import load_dataset
ds = load_dataset("openbmb/UltraFeedback")Preview a sample row
{
"instruction": "How do I make a cake?",
"completions": [
{ "response": "Mix flour, sugar...", "score": 8.5 },
{ "response": "Buy one from the store.", "score": 2.0 }
],
"correct_ranking": [0, 1]
}Fine-tune with this dataset
Estimated VRAM to fine-tune with QLoRA (4-bit base model + LoRA adapters), using conservative defaults:
New to fine-tuning? Follow the step-by-step walkthrough: Fine-Tune Your First LLM in 1 Hour
Frequently asked questions
Can I use UltraFeedback commercially?
Yes — UltraFeedback is released under MIT, 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 UltraFeedback contain, and do I need all of it?
UltraFeedback contains 64k Rows. 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 UltraFeedback best used for?
The default DPO preference set to run after any SFT pass. It belongs to the Preference (RLHF / DPO) section of our dataset hub, where you'll find alternatives and complementary sets.