File size: 702 Bytes
d1b46a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f7cb21b
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: quality
    dtype: float64
  splits:
  - name: train
    num_bytes: 1794827
    num_examples: 3805
  download_size: 1129498
  dataset_size: 1794827
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---


Qwen2 7Bで埋め込み表現を作成、LightGBMで品質スコア付け  
LightGBMのトレーニングにはneody/wikipedia-2023-11-ja-cleanedから、1. 文章の長さが上位0.001%以内のもの 2. features(スコア)の値が0.68以上のものからnegativeとして5000件、この条件に当てはまらないものの中から5000件をpositiveとしてサンプリング