指令 / SFT · HuggingFaceTB

Smoltalk

A high-quality 2025 synthetic dataset created to train the SmolLM2 family of small language models. 1 million examples covering diverse tasks, generated with careful quality filtering. Demonstrates that small models (1.7B, 360M) can achieve strong performance with the right training data.

加载
from datasets import load_dataset
ds = load_dataset("HuggingFaceTB/smoltalk")
预览示例数据
{
  "messages": [
    { "role": "user", "content": "Summarize the key differences between TCP and UDP." },
    { "role": "assistant", "content": "TCP is connection-oriented: it establishes a session, guarantees delivery, and retransmits lost packets. UDP is connectionless: it fires packets without handshaking or delivery guarantees, making it faster but less reliable. Use TCP for web traffic and file transfers; UDP for video streaming, DNS, and gaming." }
  ]
}

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 VRAM13B 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

常见问题

Smoltalk 可以商用吗?

可以——Smoltalk 采用 Apache 2.0 宽松许可证,允许商业使用,包括训练用于产品的模型。发布前请查看数据集卡片中的署名要求。

Smoltalk 有多少数据?需要全部使用吗?

Smoltalk 包含 1M 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。

Smoltalk 最适合做什么?

General SFT for small models (the SmolLM2 recipe)。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。