metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: lang
      dtype: string
    - name: type
      dtype: string
    - name: id
      dtype: string
  splits:
    - name: eval
      num_bytes: 76408631
      num_examples: 10000
  download_size: 39911840
  dataset_size: 76408631
configs:
  - config_name: default
    data_files:
      - split: eval
        path: data/eval-*
license: odc-by
language:
  - fr
  - en
  - es
tags:
  - Python
  - Java
  - JavaScript
  - C/C++

Dataset Card for `dataset-eval`

Description

The dataset-eval dataset is a multilingual and multi-domain dataset designed for evaluating language model performance during training. It can be used for performance tracking, generalization diagnostics across languages or domains, and for implementing early stopping mechanisms.

The examples included were automatically selected as High quality by the EuroBERT-210m-Quality model, trained to estimate web text quality in multiple languages.

Dataset Composition

Natural Languages:
- English: 2,640 examples (from FineWeb)
- French: 2,720 examples (from FineWeb-2)
- Spanish: 2,640 examples (from FineWeb-2)
Programming Languages (from The-Stack-v2-dedup):
- Python: 500 examples
- Java: 500 examples
- JavaScript: 500 examples
- C: 250 examples
- C++: 250 examples
Total: 10,000 high-quality examples

Data Structure

Each example includes the following fields:

text (string): the textual content or source code.
lang (string): the language of the content (e.g., English, French, Spanish, Python, C++, etc.).
type (string): the type of content:
- "NL" for natural language
- "CL" for code language
id (string): a unique identifier generated by hashing the text field.

Use Cases

This dataset is intended for periodic evaluation during language model training:

Tracking performance on high-quality data
Evaluation per batch or epoch
Validation metric computation for early stopping
Performance comparison by language or domain

It is not intended for direct training, due to its limited size and its purpose as a filtered evaluation sample.

Licenses

The dataset is built from sources under the following licenses:

Source	License
FineWeb	ODC-BY 1.0
FineWeb-2	ODC-BY 1.0
The Stack v2	Other
EuroBERT-210m-Quality	Apache-2.0

Users must ensure they comply with the specific license conditions when reusing or redistributing this data.

Risks and Limitations

Sensitive Data

The original sources are from the public web and were automatically cleaned. Despite filtering, some data may still contain sensitive, personal, or confidential information.

It is strongly recommended not to use this dataset in production or user-facing systems without manual review.

Bias

Quality annotations were produced by an automatic classifier and may reflect its training biases.
The dataset covers only three natural languages and five programming languages.
Cultural, thematic, or syntactic biases may be present depending on the source corpora.

Dataset Card for dataset-eval