README.md

metadata

dataset_info:
  features:
    - name: uuid
      dtype: string
    - name: prompt
      list:
        - name: content
          dtype: string
        - name: role
          dtype: string
    - name: completion
      list:
        - name: content
          dtype: string
        - name: role
          dtype: string
    - name: original_prompt
      list:
        - name: content
          dtype: string
        - name: role
          dtype: string
    - name: original_completion
      list:
        - name: content
          dtype: string
        - name: role
          dtype: string
    - name: generator
      dtype: string
    - name: translator
      dtype: string
  splits:
    - name: train
      num_bytes: 228737923
      num_examples: 49983
  download_size: 125975016
  dataset_size: 228737923
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-nc-4.0
language:
  - sl

Nemotron STEM Translated Dataset

Overview

This is a translated (Slovene) version of STEM split from nvidia/Nemotron-Post-Training-Dataset-v2. The dataset was first subsampled to 50k examples and then translated to Slovene using Gemini 2.5 Flash. The dataset was used during the SFT phase of GaMS3 and GaMS2 models.

Data Structure

All tasks contain the following features:

• uuid: Kept from the original dataset
• prompt: Slovene translation of the prompt.
• completion: Slovene translation of the completion.
• original_prompt: Prompt from the original dataset.
• original_completion: Completion from the original dataset.
• generator: Model used for generating the original example. Kept from the original dataset
• translator: Model used for translating the example to Slovene

Acknowledgment

The dataset was developed within the PoVeJMo research program (Adaptive Natural Language Processing with Large Language Models), particularly within the research project titled SloLLaMai -- Open-access computationally efficient models for Slovenian. The program is funded within the Recovery and Resilience Plan by the Slovenian Research and Innovation Agency (ARIS) and NextGenerationEU. The authors also acknowledge the financial support from the Slovenian Research and Innovation Agency (research core funding No. P6-0411 -- Language Resources and Technologies for Slovene).

This project is also funded by the European Union under Horizon Europe (101186647 – AI4DH).

License

The dataset is released under the CC BY-NC 4.0 license, which allows only non-commercial use of this dataset.

Citation

misc{vreš2026buildingstronginstructionlanguage,
      title={Building a Strong Instruction Language Model for a Less-Resourced Language}, 
      author={Domen Vreš and Tjaša Arčon and Timotej Petrič and Dario Vajda and Marko Robnik-Šikonja and Iztok Lebar Bajec},
      year={2026},
      eprint={2603.01691},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.01691}, 
}