SmolTalk 2 — LLM Instruction / SFT Dataset

The fully open post-training corpus behind SmolLM3-3B, organized into three subsets matching the model's training phases: mid-training (4.8M rows), SFT (a decontaminated mixture of ~24 datasets including OpenThoughts, Tulu 3, OpenHermes, and multilingual data), and preference (447k rows for APO). The successor to SmolTalk — a complete, reproducible recipe for modern small-model post-training including dual reasoning/no-reasoning modes.

Dataset Details

ProviderHuggingFaceTB
CategoryInstruction / SFT
Size3 Subsets (Mid 4.8M Rows)
LicenseApache 2.0
Downloadsn/a
TagsPost-Training, SmolLM3, Reasoning, Multilingual, 2025
from datasets import load_dataset
ds = load_dataset("HuggingFaceTB/smoltalk2")

Fine-tune with this dataset

Estimated VRAM to fine-tune with QLoRA (4-bit base model + LoRA adapters), using conservative defaults:

7B QLoRA~6GB VRAM
13B QLoRA~10GB VRAM

Check if your GPU can fine-tune this →

New to fine-tuning? Follow the step-by-step walkthrough: Fine-Tune Your First LLM in 1 Hour

Related datasets

Frequently asked questions

Can I use SmolTalk 2 commercially?
Yes — SmolTalk 2 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 SmolTalk 2 contain, and do I need all of it?
SmolTalk 2 contains 3 Subsets (Mid 4.8M 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 SmolTalk 2 best used for?
Reproducing a complete modern post-training pipeline (mid-training → SFT → preference) for small models. It belongs to the Instruction / SFT section of our dataset hub, where you'll find alternatives and complementary sets.

← All datasets | Fine-Tuning Guide