Smoltalk2: 1M-10M Text Records for SmolLM3 Post-Training

Smoltalk2 contains between 1 million and 10 million text records organized into three subsets for the post-training of the SmolLM3-3B model. Developed by HuggingFaceTB and updated in October 2025, the collection supports the Mid, SFT, and Preference alignment stages of model development.

Use Cases

• Supervised fine-tuning of small language models using the SFT subset
• Preference alignment via DPO or PPO using the Preference subset
• Intermediate model refinement using the Mid subset

Strengths

• 1M-10M record scale
• Three-stage categorization (Mid, SFT, Preference)
• Directly linked to Arxiv research papers 2410.15553 and 2412.15115

Limitations

• Specific column schema not provided in the input metadata
• Requires external training recipes from the SmolLM repository for optimal subset weighting

Provenance

Source

HuggingFaceTB

Collection Method

Curated for the SmolLM3-3B post-training pipeline

Freshness

Last updated October 31, 2025.

Geography

Global