Datasets:
File size: 2,491 Bytes
269f9ad dd00589 269f9ad |
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 |
---
license: odc-by
task_categories:
- text-generation
language:
- en
pretty_name: Primus-FineWeb
tags:
- cybersecurity
- pretraining
- FineWeb
size_categories:
- 1M<n<10M
extra_gated_fields:
Affiliation: text
Country: country
I want to use this model for:
type: select
options:
- Research
- Commercial
- label: Other
value: other
Job title:
type: select
options:
- Student
- Research graduate
- AI researcher
- AI developer/engineer
- Cybersecurity researcher
- Reporter
- Other
geo: ip_location
library_name: transformers
---
> ⭐ Please download the dataset from [here](https://huggingface.co/datasets/trendmicro-ailab/Primus-FineWeb).
# PRIMUS: A Pioneering Collection of Open-Source Datasets for Cybersecurity LLM Training
## 🤗 Primus-FineWeb
The **Primus-FineWeb** dataset is constructed by filtering cybersecurity-related text from FineWeb, a refined version of Common Crawl. We began by leveraging _Primus-Seed_, a high-quality dataset of manually curated cybersecurity text, as positive samples. We then sampled ten times the amount of data from FineWeb as negative samples and trained a **binary cybersecurity classifier** based on TinyBERT. Using this classifier, we assigned each text in FineWeb a score between **0 and 1** and filtered out texts with a score greater than **0.003**, creating the Primus-FineWeb with 15.3 billion tokens. However, after discovering a significant amount of duplicate content, we performed deduplication, reducing the final dataset to **🔥 2.57 billion tokens of cybersecurity corpus**.
🚀🚀 For more details, see our paper:
[https://arxiv.org/abs/2502.11191](https://arxiv.org/abs/2502.11191)
---
## Why was the threshold set at 0.003?
We divided the score range (0-1) into several bins and randomly sampled 50 examples from each bin. These samples were then scored by GPT-4o to determine the proportion of text that was "_truly_" cybersecurity-related. We found that if the score was below 0.003, the proportion of cybersecurity text fell below 50%.
<img src="https://i.imgur.com/XbqpmbI.png" alt="Threshold Selection" width="60%">
## FineWeb: Cybersecurity Score vs. Token Count
<img src="https://i.imgur.com/6twJL1p.png" alt="Cybersecurity Score vs. Token Count" width="65%">
---
## License
This dataset is released under the **ODC-By** license. However, you must still comply with the **FineWeb license** and the **Common Crawl Terms of Use**. |