Skip to content

Dataset · HuggingFaceTB

SmolLM3-3B post-training data

HuggingFaceTB/smoltalk2

1M to 10M text examples across three subsets for intermediate training, SFT, and preference alignment of SmolLM3.

What's inside

The dataset contains three subsets: Mid, SFT, and Preference. These correspond to the three post-training phases of the SmolLM3-3B model. The card includes links to a blog post and a training recipe with subset weights. Loaded via datasets; language is not specified in metadata, modality is text.

What you can do with it

  • SFT for SmolLM3-3B or a similar small LLM
  • DPO/preference training
  • Reproducing the SmolLM3 post-training recipe
Size
1M–10M rows
Downloads this month
61.2K
Likes
188

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("HuggingFaceTB/smoltalk2")