---
language:
- ro
dataset_info:
  features:
  - name: id
    dtype: string
  - name: text
    dtype: string
  - name: url
    dtype: string
  - name: date
    dtype: string
  - name: dump
    dtype: string
  - name: file_path
    dtype: string
  - name: language_score
    dtype: float64
  - name: minhash_cluster_size
    dtype: int64
  - name: top_langs
    dtype: string
  - name: output
    dtype: string
  - name: explanation
    dtype: string
  - name: int_score
    dtype: string
  - name: age_group
    dtype: string
  - name: topic
    dtype: string
  - name: subtopic
    dtype: string
  - name: format
    dtype: string
  splits:
  - name: train
    num_bytes: 5492801065
    num_examples: 1058200
  download_size: 2976465017
  dataset_size: 5492801065
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# FineWeb2-Ro-LLM

**FineWeb2-Ro-LLM** is a high-quality pretraining dataset for the Romanian language. The data comes from [FineWeb2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) and
annotated leveraging LLMs. More details can be found [here](https://arxiv.org/abs/2511.01090).

## Key Features

* **High Quality**: The dataset was annotated using [Gemma3 12B](https://huggingface.co/google/gemma-3-12b-it)
* **Large Scale**: Contains approximately **1.06M** documents (rows).
* **Rich Metadata**: Includes detailed metadata such as quality scores (`int_score`), topics, subtopics, and reasoning/explanations for the assigned quality scores.

## Usage

You can load this dataset using the Hugging Face `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("OpenLLM-Ro/fineweb2-ro-llm", split="train")