README.md

metadata

dataset_info:
  features:
    - name: id
      dtype: string
    - name: url
      dtype: string
    - name: title
      dtype: string
    - name: chunks
      sequence: string
    - name: embeddings
      sequence:
        sequence: float32
  splits:
    - name: train
      num_bytes: 5021489124
      num_examples: 534044
  download_size: 4750515911
  dataset_size: 5021489124
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
language:
  - cs
size_categories:
  - 100K<n<1M
task_categories:
  - text-generation
  - fill-mask
license:
  - cc-by-sa-3.0
  - gfdl

This dataset contains the Czech subset of the wikimedia/wikipedia dataset. Each page is divided into paragraphs, stored as a list in the chunks column. For every paragraph, embeddings are created using the intfloat/multilingual-e5-base model.

Usage

Load the dataset:

from datasets import load_dataset

ds = load_dataset("karmiq/wikipedia-embeddings-cs-e5-base", split="train")
ds[1]

{
  'id': '1',
  'url': 'https://cs.wikipedia.org/wiki/Astronomie',
  'title': 'Astronomie',
  'chunks': [
    'Astronomie, řecky αστρονομία z άστρον ( astron ) hvězda a νόμος ( nomos )...',
    'Myšlenky Aristotelovy rozvinul ve 2. století našeho letopočtu Klaudios Ptolemaios...',
    ...,
  ],
  'embeddings': [
    [0.09006806463003159, -0.009814552962779999, ...],
    [0.10767366737127304, ...],
    ...
  ]
}

The structure makes it easy to use the dataset for implementing semantic search.

Load the data in Elasticsearch

def doc_generator(data, batch_size=1000):
  for batch in data.with_format("numpy").iter(batch_size):
    for i, id in enumerate(batch["id"]):
      output = {"id": id}
      output["title"] = batch["title"][i]
      output["url"] = batch["url"][i]
      output["parts"] = [
          { "chunk": chunk, "embedding": embedding }
          for chunk, embedding in zip(batch["chunks"][i], batch["embeddings"][i])
      ]
      yield output

num_indexed, num_failed = 0, 0,
progress = tqdm(total=ds.num_rows, unit="doc", desc="Indexing")

for ok, info in parallel_bulk(
    es,
    index="wikipedia-search",
    actions=doc_generator(ds),
    raise_on_error=False,
):
    if not ok:
        print(f"ERROR {info['index']['status']}: "
              f"{info['index']['error']['type']}: {info['index']['error']['caused_by']['type']}: "
              f"{info['index']['error']['caused_by']['reason'][:250]}")

    progress.update(1)

Use sentence_transformers.util.semantic_search

import sentence_transformers
model = sentence_transformers.SentenceTransformer("intfloat/multilingual-e5-base")

ds.set_format(type="torch", columns=["embeddings"], output_all_columns=True)

# Flatten the dataset
def explode_sequence(batch):
  output = { "id": [], "url": [], "title": [], "chunk": [], "embedding": [] }

  for id, url, title, chunks, embeddings in zip(
    batch["id"], batch["url"], batch["title"], batch["chunks"], batch["embeddings"]
  ):
    output["id"].extend([id for _ in range(len(chunks))])
    output["url"].extend([url for _ in range(len(chunks))])
    output["title"].extend([title for _ in range(len(chunks))])
    output["chunk"].extend(chunks)
    output["embedding"].extend(embeddings)

  return output

ds_flat = ds.map(
  explode_sequence,
  batched=True,
  remove_columns=ds.column_names,
  num_proc=min(os.cpu_count(), 32),
  desc="Flatten")
ds_flat

query = "Čím se zabývá fyzika?"

hits = sentence_transformers.util.semantic_search(
  query_embeddings=model.encode(query),
  corpus_embeddings=ds_flat["embedding"],
  top_k=10)

for hit in hits[0]:
    title = ds_flat[hit['corpus_id']]['title']
    chunk = ds_flat[hit['corpus_id']]['chunk']
    print(f"[{hit['score']:0.2f}] {textwrap.shorten(chunk, width=100, placeholder='…')} [{title}]")

# [0.90] Fyzika částic ( též částicová fyzika ) je oblast fyziky, která se zabývá částicemi. V širším smyslu… [Fyzika částic]
# [0.89] Fyzika ( z řeckého φυσικός ( fysikos ): přírodní, ze základu φύσις ( fysis ): příroda, archaicky… [Fyzika]
# ...

The embeddings generation took about 2 hours on an NVIDIA A100 80GB GPU.

License

See license of the original dataset: https://huggingface.co/datasets/wikimedia/wikipedia.