aisuko/simple_english_wikipedia

Only for the reaseaching usage.

The original data from http://sbert.net/datasets/simplewiki-2020-11-01.jsonl.gz.

We use nq_distilbert-base-v1 model encode all the data to the PyTorch Tensors. And normalize the embeddings by using sentence_transformers.util.normalize_embeddings.

How to use

See notebook Wikipedia Q&A Retrieval-Semantic Search

Installing the package

!pip install sentence-transformers==2.3.1

The converting process

# the whole process takes 1287.0s GPU P100

import os
import json
import gzip
from sentence_tranformers.util import http_get
from sentence_transformers import SentenceTransformer
from sentence_transformers.util import normalize_embeddings


os.environ['DATASET_NAME']='simplewiki-2020-11-01.jsonl.gz'
os.environ['DATASET_URL']='http://sbert.net/datasets/simplewiki-2020-11-01.jsonl.gz'

os.environ['MODEL_NAME']='multi-qa-MiniLM-L6-cos-v1'
os.environ['CROSS_CODE_NAME']='cross-encoder/ms-marco-MiniLM-L-6-v2'


http_get(os.getenv('DATASET_URL'), os.getenv('DATASET_NAME'))

passages=[]
with gzip.open(os.getenv('DATASET_NAME'), 'rt', encoding='utf-8') as fIn:
    for line in fIn:
        data=json.loads(line.strip())
        # add all paragraphs
#         passages.extend(data['paragraphs'])

        # only add the first paragraph
#         passages.append(data['paragraph'][0])

        for paragraph in data['paragraphs']:
            # We encode the passages as [title, text]
            passages.append([data['title'], paragraph])

print('Passages:', len(passages))      


bi_encoder=SentenceTransformer('nq-distilbert-base-v1')
bi_encoder.max_seq_length=256
bi_encoder.to('cuda')

corpus_embeddings=bi_encoder.encode(passages, convert_to_tensor=True, show_progress_bar=True).to('cuda')
corpus_embeddings=normalize_embeddings(corpus_embeddings)
len(corpus_embeddings)


import pandas as pd

embedding_data=pd.DataFrame(corpus_embeddings.cpu())
embedding_data.to_csv('simple_english_wikipedia_2020_11_01.csv', index=False)