File size: 4,497 Bytes
661cf56
 
 
 
71035b7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
795d2a6
 
 
 
71035b7
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
---
license: apache-2.0
language:
- en
task_categories:
- text-generation
tags:
- llm
- pretraining
- web
- data-selection
size_categories:
- n>1T
---

# FineWeb-Mask


[πŸ“œ DATAMASK Paper](https://arxiv.org/abs/2512.24265) | [πŸ’» GitHub Repository](https://github.com/ByteDance-Seed/DATAMASK) | [πŸ“¦ Fineweb-Mask Dataset](https://huggingface.co/datasets/DATA-MASK/FineWeb-Mask)

</div>

## πŸ“š Introduction

**FineWeb-Mask** is a 1.5 trillion token, high-efficiency pre-training dataset curated using the **DATAMASK** framework. Developed by the **ByteDance Seed team**, DATAMASK addresses the fundamental tension in large-scale data selection: the trade-off between **high quality** and **high diversity**.

By modeling data selection as a **Mask Learning** problem, we provide a derivative of the original [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) corpus. FineWeb-Mask is designed to eliminate semantic redundancy while preserving the highest quality samples, allowing models to achieve superior performance with significantly less data.

## 🎯 The Problem: The Quality-Diversity Trap

In large language model (LLM) pre-training, developers usually face two suboptimal choices:

1. **The Quality Trap:** Filtering solely by quality scores leads to "diminishing returns." Samples become highly clustered, resulting in severe semantic redundancy.
2. **The Diversity Trap:** Filtering solely for diversity often discards high-value quality samples, leading to worse performance than the original raw dataset.
3. **The Compute Bottleneck:** Traditional diversity algorithms (like greedy selection) are computationally prohibitive for trillion-token datasets.

## πŸ’‘ Highlights: The DATAMASK Framework

DATAMASK breaks this deadlock through a "joint harvesting" strategy:

* **Joint Optimization:** Uses Policy Gradient algorithms to optimize both quality and diversity metrics within a unified framework.
* **Extreme Acceleration:** Through probability relaxation and specialized optimization techniques, DATAMASK reduces computation time by **98.9%** compared to traditional greedy algorithms, making trillion-token selection feasible.
* **The "Balancer":** Includes a tunable parameter  that allows developers to define the "Golden Ratio" between quality and diversity for their specific needs.
* **Semantic De-redundancy:** Visual analysis shows that FineWeb-Mask samples are distributed evenly across high-quality regions rather than being rigidly clustered.

## πŸ“ˆ Evaluation Results

FineWeb-Mask demonstrates that **1+1 > 2**. By selecting a subset that represents only ~10% of the original scale in specific experiments, we observed:

* **Dense Models:** A **3.2% average improvement** across 12 benchmarks for 1.5B dense models.
* **MoE Models:** A **1.9% improvement** for 7B Mixture-of-Experts (MoE) models.
* **Length Bias Correction:** While quality filters favor long text and diversity filters favor short text, DATAMASK finds a scientific middle ground.

| Model Size | Dataset | Avg. Score (12 Benchmarks) | Improvement |
| --- | --- | --- | --- |
| 1.5B Dense | FineWeb (Original) | Baseline | - |
| 1.5B Dense | **FineWeb-Mask** | **+3.2%** | πŸš€ |
| 7B MoE | FineWeb (Original) | Baseline | - |
| 7B MoE | **FineWeb-Mask** | **+1.9%** | πŸš€ |

## ❀️ Acknowledgements

FineWeb-Mask is built upon the incredible foundational work of the [HuggingFace FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) team. We are grateful to the open-source community for providing the raw corpora that made this optimization possible.

## 🌟 Citation

If you find our dataset or the DATAMASK framework useful, please cite our work:

```bibtex
@misc{fan2025jointselectionlargescalepretraining,
      title={Joint Selection for Large-Scale Pre-Training Data via Policy Gradient-based Mask Learning}, 
      author={Ziqing Fan and Yuqiao Xian and Yan Sun and Li Shen},
      year={2025},
      eprint={2512.24265},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.24265}, 
}

```

## πŸ’³ License

This dataset is released under the **Apache 2.0** license. Users should also adhere to the original license terms of the FineWeb dataset and its constituent sources.

## **πŸ“§** Contact
- Ziqing Fan: zqfan_knight@sjtu.edu.cn
- Yuqiao Xian: ericxian1997@gmail.com

---

**Would you like me to help you draft the "How to Use" section for loading this dataset via the Hugging Face `datasets` library?**