ner-collection

A local dataset of named-entity recognition corpora, converted to Parquet for NER research.

52 source entries consist of 16,925,066 records in 294 Parquet files, grouped into 163 configs. Every record points back to its original file, kept verbatim in raw/<corpus>.tar.gz. manifest.json holds the config list, per-file SHA-256 checksums, and the schema of every column.

707 rows have kind: "invalid": the source data itself is broken (353 + 353 null annotations in ontonotes/ontovalidate, one malformed JSON line in wikineural).

Run

import json
from datasets import load_dataset

configs = json.load(open("manifest.json", encoding="utf-8"))["configs"]
by_name = {c["config_name"]: c for c in configs}

cfg = by_name["RuNNE__json"]
files = {d["split"]: d["path"] for d in cfg["data_files"]}
ds = load_dataset("parquet", data_files=files)

row = ds["train"][0]
print(row["kind"], row["text"][:80])
print(json.loads(row["entities"])[0])

kind says what shape a row has: token_sequence fills tokens + ner_tags (BIO), while sentence_spans / document_spans fill text + entities (character offsets). Swap RuNNE__json for any other config name from manifest.json → configs.

To read an original back:

import tarfile
with tarfile.open(row["asset_archive"]) as tf:
    original = tf.extractfile(row["asset_member"]).read()