tanganke/smol-smoltalk-Interaction-SFT-formatted

Formatted version in standard conversation format for SFT. Below is the original README.

Dataset Card for ReactiveAI/Smol-Smoltalk Interaction SFT

Derived from HuggingFaceTB/smol-smoltalk. Made for Interaction Supervised Fine-Tuning of Reactive Transformer Proof-of-Concept models, especially RxT-Beta.

Dataset Details

Dataset Description

Reactive Transformers are processing only the single interactions in real-time and using Short-Term Memory to store information from previous interactions. Before the model is able to use it's memory, it has to be fine-tuned to its interaction format - question/query and anwer/response.

Original Dataset has messages in JSON format with full conversations and messages with different roles. Reactive models aren't using system prompt (their initial memory states could take role of system prompt), so we are not including system prompts from original Dataset. Every pair of consecutive user and assistant messages are added to new Dataset's query and response fields. Each interaction is treated as separate one, because on SFT stage, models are still not using memory.

Original HuggingFace Dataset Description

This is a subset of SmolTalk dataset adapted for smol models with less than 1B parameters. We used it to build SmolLM2-360M-Instruct and SmolLM2-135M-Instruct. We do SFT on this dataset and then DPO on UltraFeedback.

Compared to SmolTalk:

• The conversations from Smol-Magpie-Ultra are shorter in this dataset

• We include less task specific data compared to SmolTalk (e.g no function calling and less rewriting and summarization examples) since these smaller models have limited capacity

• We don't include any advanced math datasets

• Language(s) (NLP): English only

• License: Apache-2.0 originally licensed for Loubna Ben Allal and Anton Lozhkov and Elie Bakouch and Gabriel Martín Blázquez and Guilherme Penedo and Lewis Tunstall and Andrés Marafioti and Hynek Kydlíček and Agustín Piqueres Lajarín and Vaibhav Srivastav and Joshua Lochner and Caleb Fahlgren and Xuan-Son Nguyen and Clémentine Fourrier and Ben Burtenshaw and Hugo Larcher and Haojun Zhao and Cyril Zakka and Mathieu Morlon and Colin Raffel and Leandro von Werra and Thomas Wolf

Dataset Sources

• Repository: HuggingFaceTB/smol-smoltalk
• Extended Version: HuggingFaceTB/smoltalk

Uses

This dataset is made for Interaction Supervised Fine-Tuning for reactive models. Reactive models are processing only single sequences (interactions). Dataset contains records with query and answer fields, that look like they are taken from the conversation, but they are not inter-connected.

Additionally, each row includes source field as in original dataset - Dataset composition

Direct Use

This dataset should be used to train small scale research models.

Out-of-Scope Use

Dataset is rather too small to be used for production models

Dataset Structure

[More Information Needed]

Dataset Creation

Curation Rationale

[More Information Needed]

Recommendations

Please check original Dataset docs for more detailed information about content - Dataset Sources

Citation

From original HuggingFace dataset

@misc{allal2025smollm2smolgoesbig,
      title={SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model}, 
      author={Loubna Ben Allal and Anton Lozhkov and Elie Bakouch and Gabriel Martín Blázquez and Guilherme Penedo and Lewis Tunstall and Andrés Marafioti and Hynek Kydlíček and Agustín Piqueres Lajarín and Vaibhav Srivastav and Joshua Lochner and Caleb Fahlgren and Xuan-Son Nguyen and Clémentine Fourrier and Ben Burtenshaw and Hugo Larcher and Haojun Zhao and Cyril Zakka and Mathieu Morlon and Colin Raffel and Leandro von Werra and Thomas Wolf},
      year={2025},
      eprint={2502.02737},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.02737}, 
}