---
dataset_info:
  features:
  - name: conversations
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  - name: original
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  splits:
  - name: train
    num_bytes: 4076499503
    num_examples: 958729
  download_size: 1787731931
  dataset_size: 4076499503
datasets:
  - NousResearch/Hermes-3-Dataset
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
language:
  - uk
pretty_name: Hermes 3 Ukrainian Fixed Conversations
tags:
  - ukrainian
  - llm
  - nlp
license: cc-by-sa-4.0
task_categories:
  - text-generation
  - question-answering
  - summarization
---

# Dataset Card for Hermes 3 Ukrainian Fixed Conversations

## Dataset Description

**Dataset Summary**

**hermes3-uk-fixed** is a Ukrainian translation of the **[NousResearch/Hermes-3-Dataset]**. The translation was produced with *Gemma 3 27B (instruction-tuned)*. During preparation we **removed all system prompts** and **normalized the message roles and content** to match the common schema we use across our dialog datasets.

**Languages**

- Ukrainian (uk)

## Dataset Structure

**Data Fields**

- `conversations`: list of messages in a dialog (array of objects)
  - `from`: normalized sender role — **`user`** or **`assistant`** (system messages are removed)
  - `value`: message text

**Splits**

- `train`: full set of dialogs (count matches the source dataset after filtering/merging).

## Dataset Creation

**Source Data**

- Base dataset: *NousResearch/Hermes-3-Dataset* (Apache-2.0).
- Translation: inference with *google/gemma-3-27b-it*.

**Processing / “Fixed” Changes**

1. **Removed `system`**: dropped all system-level messages from `conversations`.
2. **Role normalization**: mapped `human` → `user`, `gpt` → `assistant`; collapsed any aliases to these two roles.

**Intended Uses**

- Instruction/chat LLM training in Ukrainian
- Continued pretraining or SFT on multi-turn dialogs
- Research on adapting Hermes-style dialogs to Ukrainian

## Considerations for Using the Data

**Social Impact**

Aims to strengthen the Ukrainian-language LLM ecosystem and improve accessibility of language technology for Ukrainian speakers.

**Bias & Limitations**

- Hermes-3 includes neutral and roleplay/creative dialogs — stylistic bias may be present.
- Machine translation can introduce subtle errors or tone shifts.
- Removing `system` prompts may drop context that affected certain responses.

**Recommendations**

- Suitable for general dialog fine-tuning and instruction following.
- For specialized domains, apply additional filtering and QA.

## How to Use

```python
from datasets import load_dataset

ds = load_dataset("le-llm/hermes3-uk-fixed", split="train")
print(ds[0]["conversations"][:2])
# [{'from': 'user', 'value': '...'}, {'from': 'assistant', 'value': '...'}]
```


## Citation

When using this dataset, please cite the original sources:

- Nous Research. *Hermes-3-Dataset* (Apache-2.0), Hugging Face Datasets.
- Google DeepMind. *Gemma 3 27B (IT)*, Hugging Face Models / Gemma 3 documentation.

**BibTeX**

TBD, pleliminary below


```bibtex
@dataset{hermes3_uk_fixed_2025,
  title   = {Hermes 3 Ukrainian Fixed Conversations},
  author  = {le-llm},
  year    = {2025},
  url     = {https://huggingface.co/datasets/lapa-llm/hermes3-uk-fixed}
}

@misc{nous_hermes3_dataset,
  title        = {Hermes-3-Dataset},
  author       = {Nous Research},
  howpublished = {Hugging Face Datasets},
  year         = {2024},
  url          = {https://huggingface.co/datasets/NousResearch/Hermes-3-Dataset}
}

@misc{gemma3_27b_it_2025,
  title        = {Gemma 3 27B (instruction-tuned)},
  author       = {Google DeepMind},
  howpublished = {Hugging Face Models},
  year         = {2025},
  url          = {https://huggingface.co/google/gemma-3-27b-it}
}
```

## License

CC-BY-SA-4.0

---

*This dataset is part of the "Lapa" - Ukrainian LLM initiative to advance natural language processing for the Ukrainian language.*