Datasets:
Multi-SmolTalk-Summarize (no think)
Multi-SmolTalk-Summarize (no think) is the multilingual, re-generated version of Smol-summarize, the summarization subset of SmolTalk. It is released in English, Italian, French, German and Spanish and contains about 96,000 summarization samples in total, distributed with 40% English and 15% for each of the other four languages.
The dataset is a supervised fine-tuning (SFT) resource for the summarization task: each sample is a three-turn chat in which a system prompt defines the desired output format and constraints, the user turn contains an email or a news article, and the assistant turn contains the summary.
Source dataset
| Dataset | Smol-summarize (configuration smol-summarize of SmolTalk) |
| Hugging Face | HuggingFaceTB/smoltalk |
| Paper | SmolLM2: When Smol Goes Big — Data-Centric Training of a Small Language Model (Ben Allal et al., 2025) |
| License | Apache 2.0 |
Smol-summarize is a 100k-sample synthetic collection built with distilabel. The source documents (emails and news articles) were generated with Qwen2.5-72B-Instruct conditioned on personas from PersonaHub and FinePersonas, providing a variety of topics, lengths and registers; the same model was then prompted to summarize them. Each sample is a single conversation with a system prompt that specifies the summarization instruction, an email or news article, and the reference summary. The release provides train (96,356 rows) and test (5,072 rows) splits, with a single messages column in chat format.
How the dataset was built
Two changes were applied to the source data: the instruction set was expanded, and the summaries were regenerated so that every output is consistent with its new instruction. In addition, the source samples were divided into five disjoint language shares: 40% of the documents were kept in English and 15% were assigned to each of the other four languages.
- Expanded instructions: the original list of system prompts was extended with 12 additional custom instructions, which add formal constraints (maximum number of sentences, words or paragraphs, bulleted and numbered lists, tables, concept maps, extractive summaries) and content constraints (key concepts, actions and intentions, people and organisations, dates and locations).
- New summaries: each source document was paired with an instruction from the expanded list and the corresponding summary was generated again.
- English configuration: for the English share, the summaries were regenerated starting from the original emails and news articles.
- Italian, French, German and Spanish configurations: for each of the other shares, the original email or news article was first translated into the target language through an automated translation pipeline, and the summary was then generated directly in that language from the translated source.
- Models: the translation and regeneration were performed with
RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamicandQwen/Qwen3-VL-235B-A22B-Instruct-FP8.
Dataset structure
Configurations
| Config | Language | Entries | Share |
|---|---|---|---|
en |
English | 38,424 | 40% |
fr |
French | 14,410 | 15% |
de |
German | 14,409 | 15% |
it |
Italian | 14,409 | 15% |
es |
Spanish | 14,402 | 15% |
| Total | 96,054 | 100% |
Fields
| Field | Description |
|---|---|
messages |
List of chat messages: a user turn with the email or news article and an assistant turn with the generated summary |
chat_template_kwargs |
Per-sample options for the chat template: custom_instructions (including the summarization request), enable_thinking, python_tools and xml_tools |
source |
Provenance identifier of the source Smol-summarize sample the row was derived from |
translation_model |
Name of the model used to translate the document into the target language |
generation_model |
Name of the model used to generate the summary |
Usage
from datasets import load_dataset
ds = load_dataset("VillanovaAI/multi_smoltalk_summarize_no_think", "it", split="train")
sample = ds[0]
document, summary = sample["messages"]
print(document["content"]) # email or news article
print(summary["content"]) # reference summary
Limitations
- The summaries are model-generated from synthetic documents, so they may contain factual inconsistencies or stylistic artefacts; the translated configurations add any error introduced by the translation step.
- The instruction set covers a limited number of output formats and length constraints.
Licensing and attribution
The upstream subset is released under Apache 2.0 and this multilingual version is distributed under the same license. Content and pipeline are credited to the SmolTalk / SmolLM2 team at Hugging Face.
Citation
If you use this dataset, please cite the Multi-SFT paper, which describes the whole SFT collection used to train the Villanova model family, together with the SmolLM2 paper and the original SmolTalk release.
@misc{baiamonte2026multisft,
title={Multi-SFT: Permissively-Licensed Multilingual Instruction-Following},
author={Baiamonte, Daniela and Zeinalipour, Kamyar and Fano, Elena and Simonazzi, Stefano and Zugarini, Andrea},
year={2026},
eprint={2610.22086},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2610.22086}
}
@misc{allal2025smollm2,
title={SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model},
author={Ben Allal, Loubna and Lozhkov, Anton and Bakouch, Elie and Mart{\'i}n Bl{\'a}zquez, Gabriel and Penedo, Guilherme and Tunstall, Lewis and others},
year={2025},
eprint={2502.02737},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2502.02737}
}
- Downloads last month
- 51