---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: id
    dtype: string
  - name: metadata
    struct:
    - name: date
      dtype: timestamp[us]
    - name: dump
      dtype: string
    - name: file_path
      dtype: string
    - name: int_score
      dtype: int64
    - name: language
      dtype: string
    - name: language_score
      dtype: float64
    - name: score
      dtype: float64
    - name: token_count
      dtype: int64
    - name: url
      dtype: string
  splits:
  - name: train
    num_bytes: 5292938151.266562
    num_examples: 999245
  download_size: 2716629909
  dataset_size: 5292938151.266562
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

## Qwark Corpus

1.3B+ high quality tokens from the internet, based on [HuggingFaceTB/smollm-corpus](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus)'s `fineweb-edu-dedup` subset as well as [FineMath-4+](HuggingFaceTB/finemath).

Filtering process:

| Step                                                   | Description                                            | Rows |
|--------------------------------------------------------|--------------------------------------------------------|----------------|
| 1. Stream dataset until 600K samples have been selected from SmolLM Corpus  | Keep only items with score >= 3.5             | 600,000      |
| 2. Remove items with length > 50,000                   | Filter items exceeding 50,000 characters in length     | 597,142        |
| 3. Combine with a selection of 4,000 TED transcripts   | Add educational TED talk transcripts to the dataset    | 601,147      |
| 4. Stream 400K samples from FineMath-4+ | Keep only items with score >= 4.0       | 1,001,147      |
| 5. Remove items with length > 50,000 | Filter items exceeding 50,000 characters in length     | 999,245|