LIMA: Less Is More for Alignment
Landmark alignment research showing that just 1,000 carefully curated examples rival GPT-4 in instruction-following quality. Challenges the 'more data is always better' assumption. LIMA 65B was competitive with GPT-4 on human preference tests despite only 1k training examples.
from datasets import load_dataset
ds = load_dataset("GAIR/lima")Preview a sample row
{
"conversations": [
{ "role": "user", "content": "What's the best way to learn a new programming language?" },
{ "role": "assistant", "content": "The fastest path is building something you actually care about. Pick a small project — a CLI tool, a web scraper, a game — and implement it in the new language. You'll naturally encounter the language's idioms and standard libraries through necessity rather than abstract tutorials." }
]
}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 LIMA: Less Is More for Alignment commercially?
Not in a product — LIMA: Less Is More for Alignment is released under CC-BY-NC-SA 4.0, which restricts use to research and other non-commercial purposes. For commercial fine-tuning, pick a permissively licensed dataset from the same category instead.
How much data does LIMA: Less Is More for Alignment contain, and do I need all of it?
LIMA: Less Is More for Alignment contains 1k 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 LIMA: Less Is More for Alignment best used for?
Style and format alignment with a tiny curated set - quality over quantity. It belongs to the Instruction / SFT section of our dataset hub, where you'll find alternatives and complementary sets.