LIMA: Less Is More for Alignment — LLM 指令 / SFT Dataset
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.
Dataset Details
| Provider | Meta / GAIR |
| Category | 指令 / SFT |
| Size | 1k Rows |
| License | CC-BY-NC-SA 4.0 |
| Downloads | 300k |
| Tags | Curated, Quality-over-Quantity, Research, Alignment |
from datasets import load_dataset
ds = load_dataset("GAIR/lima")
用这个数据集微调
使用 QLoRA（4-bit 基础模型 + LoRA 适配器）微调的预计显存需求（保守默认参数）：
| 7B QLoRA | ~6GB VRAM |
| 13B QLoRA | ~10GB VRAM |
检测你的 GPU 能否微调 →
微调新手？跟着分步教程走： 一小时微调你的第一个 LLM
相关数据集
- s1K-1.1 — Cheap, fast reasoning fine-tunes — 1k samples means minutes of training, not days
- No Robots — Your first fine-tune — small, clean, 100% human-written SFT data (non-commercial license)
- Python-Edu — Continued pretraining for Python code understanding
- OpenHermes 2.5 — The default general-purpose SFT mix for 7B-13B fine-tunes
常见问题
LIMA: Less Is More for Alignment 可以商用吗？
不能用于产品——LIMA: Less Is More for Alignment 采用 CC-BY-NC-SA 4.0，仅限研究等非商业用途。商业微调请改用同类别中宽松许可的数据集。
LIMA: Less Is More for Alignment 有多少数据？需要全部使用吗？
LIMA: Less Is More for Alignment 包含 1k Rows。通常不需要全部：风格和格式微调只需几百到几千条样本——先加载切片（如 split="train[:1000]"），质量到达瓶颈时再扩大规模。
LIMA: Less Is More for Alignment 最适合做什么？
Style and format alignment with a tiny curated set - quality over quantity。它属于数据集中心的「指令 / SFT」板块，那里有替代和互补的数据集。
← 全部数据集 | Fine-Tuning Guide