quality-pro / README.md
qikp's picture
Update README.md
2534004 verified
|
Raw History Blame Contribute Delete
1.27 kB
---
dataset_info:
features:
- name: text
dtype: string
- name: label
dtype: float64
- name: source
dtype: string
- name: sha256
dtype: string
splits:
- name: train
num_bytes: 11221869
num_examples: 2242
download_size: 6141002
dataset_size: 11221869
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Quality Pro
Quality Pro is a dataset designed to train educational or quality classifiers for web content.
## Sources
### bespoke_qikp
An English test split of agentlans/common-crawl-sample and ClimbMix was manually annotated by me. Subsets of both were used, with 80% Common Crawl and 20% ClimbMix.
### ultrafineweb-filtered
The pre-filtered data in Ultra-FineWeb L2. High quality scores but only 0.5-1 was included.
### c4-laya
Uncleaned C4 filtered with the Laya zero-shot classifier using this object, with each value subtracted by 1:
```json
{
"quality": {
"type": "noul",
"instructions": "Is this document spam, junk, repetitive, trivial, or minimal content?"
}
}
```
### fineweb-edu-classifier-training-data
A subset of 100 Llama 3-generated annotations used for training the official FineWeb-Edu classifier, with entries that are improperly annotated deleted.