Back to Gap Map Model components / Training & synthetic datasets

SmolTalk

open / Overall score: 4.0(strong)

Hugging Face's supervised fine-tuning mixture behind the SmolLM instruct models, combining new synthetic sets such as Smol-Magpie-Ultra with established public datasets. Marin 8B Instruct also trains on it. The record covers both SmolTalk and its successor SmolTalk2.

Component datasets keep the terms of their own sources, which the card documents.

Openness

4 high confidence

4.0

access: ungated
license: apache-2.0(new-subsets)+per-component
dataset_card: present

Ungated; newly created subsets are Apache-2.0, incorporated public datasets keep their own licenses.

https://huggingface.co/api/datasets/HuggingFaceTB/smoltalk recorded 2026-09-29
Hub API record for HuggingFaceTB/smoltalk has the gated field set to false and private set to false, and its description reads "This is a synthetic dataset designed for supervised finetuning (SFT) of LLMs".
https://huggingface.co/datasets/HuggingFaceTB/smoltalk recorded 2026-09-29
Dataset page renders the card, including "This is a synthetic dataset designed for supervised finetuning (SFT) of LLMs".
https://huggingface.co/datasets/HuggingFaceTB/smoltalk/raw/main/README.md recorded 2026-09-29
Raw dataset card for SmolTalk states "All the new datasets (Smol-Magpie-Ultra, Smol-contraints, Smol-rewrite, Smol-summarize) are licensed under" Apache 2.0 and "please refer to the original dataset for the license" for the existing public datasets, and has a "Dataset composition" section.

Adoption

3 high confidence

3.0

Adoption is measured by Hugging Face downloads of HuggingFaceTB/smoltalk, the only usage channel on record for SmolTalk.

https://huggingface.co/api/datasets/HuggingFaceTB/smoltalk recorded 2026-08-13
30,281 downloads in the trailing 30 days for HuggingFaceTB/smoltalk

Capability

5 high confidence

5.0

In the SmolLM2 paper's comparison, fine-tuning on SmolTalk beats OpenHermes-2.5, UltraChat and MagPie-Pro, and the mixture trains the SmolLM2 instruct models and is one of the supervised fine-tuning sets in Marin 8B Instruct. Its successor SmolTalk2 carries SmolLM3's post-training.

https://arxiv.org/html/2502.02737v1 recorded 2026-09-29
SmolLM2 paper section 5.1 says "Our dataset outperforms MagPie-Pro on most benchmarks, and largely surpasses OpenHermes2.5 and UltraChat" on IFEval and MT-Bench.
https://huggingface.co/datasets/HuggingFaceTB/smoltalk2/raw/main/README.md recorded 2026-09-29
SmolTalk2 card says its subsets "correspond to the three phases of Post-Training for" SmolLM3-3B.
https://huggingface.co/marin-community/marin-8b-instruct recorded 2026-09-29
Marin 8B Instruct model page lists "HuggingFaceTB/smoltalk" among its SFT datasets, alongside "allenai/tulu-3-sft-mixture".

Verified 2026-09-27