arabovs-ai-lab/TatCorp_222M

TatCorp-222M (Private)

TatCorp-222M — the largest open corpus of Tatar language texts prepared for research in Natural Language Processing (NLP) and language modelling. The corpus is provided in Parquet/JSONL format and optimized for streaming processing and training large language models.

---

📥 Access and License

• Repository: arabovs-ai-lab/TatCorp_222M (private on Hugging Face Hub)
• Access: available upon request for academic and non-commercial research after signing a Data Use Agreement (DUA).
• Redistribution: redistribution of source texts is prohibited without explicit permission from the rights holders.
• Contact: cool.araby@gmail.com

---

🗂 Structure

• Format: JSONL (one record per line) and sharded Parquet
• Total records: 541,543
• Splits: train (530,712 / ~98%), validation (5,415 / ~1%), test (5,416 / ~1%)
• Fields: id, url, title, content, category, date, lang, source, license

---

📊 Detailed Statistics

Key Metrics

• Total volume: 541,543 documents
• Tokens (text): ~222,312,958
• Tokens (titles): ~1,680,896
• Total tokens: ~223,993,854
• Average document length: 410.6 tokens
• Unique vocabulary: ~2,200,895 words
• Categories: 482 unique categories

Data Completeness

Metric	Count	Percentage
Records with URL	541,543	100.0%
Records with category	541,193	99.94%
Records with title	541,087	99.92%
Records with text	541,472	99.99%
Records with date	90,475	16.7%

Distribution by Source (TOP-10)

#	Domain	Documents	Share	Content Type
1	wikipedia.org	456,070	84.2%	Encyclopaedic
2	matbugat.ru	44,880	8.3%	News
3	intertat.tatar	19,529	3.6%	News/Cultural
4	azatliq.org	8,180	1.5%	News
5	vk.com	6,796	1.3%	Social network
6	shahrikazan.ru	2,399	0.4%	Regional
7	tatar-inform.tatar	1,524	0.3%	News
8	mamadysh-rt.ru	1,212	0.2%	Regional
9	belgech.ru	833	0.2%	Blog
10	vatantat.ru	119	0.02%	Cultural

Total unique domains: 11

Distribution by Categories (TOP-15)

#	Category	Documents	Share
1	Wikipedia	456,071	84.3%
2	Matbugat.ru	11,886	2.2%
3	Economics	6,488	1.2%
4	Shähri Kazan	6,276	1.2%
5	Vatanım Tatarstan	6,270	1.2%
6	Yañalıqlar (News)	4,191	0.8%
7	Ömet (Hope)	2,128	0.4%
8	Intertat.ru	1,991	0.4%
9	Culture	1,975	0.4%
10	Tatarstan Youth	1,591	0.3%
11	Fän (Science)	1,333	0.2%
12	Sälämätlek (Health)	1,193	0.2%
13	Sport	1,086	0.2%
14	Mädäniät (Culture)	898	0.2%
15	Tormış (Life)	895	0.2%

Total unique categories: 482

Document Length Statistics

Metric	Titles (title)	Texts (content)
Total tokens	1,680,896	222,312,958
Mean	3.1	410.6
Median	3	~250
Minimum	0	1
Maximum	26	608,353
Standard deviation	2.6	3,301.9
Unique words	344,151	2,196,225

Temporal Distribution

• Dated documents: 90,475 (16.7% of the corpus)
• Range: 2007-2025
• Majority: 2020-2025 (over 50,000 documents)
• Most recent year: 2025 (14,213 documents)

🎯 Key Insights

1. Wikipedia dominance - 84.2% of the corpus consists of encyclopaedic articles
2. Content currency - most dated documents are from 2020-2025
3. Large length variance - standard deviation of 3,301.9 indicates very long documents
4. Rich vocabulary - 2.2 million unique tokens

---

⚖️ Ethics and PII

• Source: publicly available web content
• PII removal: automatic and selective manual removal of personal data applied. Complete removal is not guaranteed — additional verification is recommended for sensitive use cases
• Collection scripts: not published for legal/ethical reasons. Post-processing code (normalisation, deduplication, filtering) is available

---

🔗 Models Trained on This Corpus

• TatarTokenizers — Trained tokeniser models and BPE/WordPiece configurations based on this corpus:
  https://huggingface.co/arabovs-ai-lab/TatarTokenizers

• Tatar2Vec — Trained word and document embedding models based on this corpus:
  https://huggingface.co/arabovs-ai-lab/Tatar2Vec

---

🔬 Potential Use Cases

The TatCorp-222M corpus can be used for the following research tasks:

1. Language modelling — fine-tuning or training large language models for Tatar from scratch
2. Text classification — multi-class categorisation by topic, genre, or source
3. Topic modelling — discovering latent topics and their temporal dynamics
4. Text generation — training models for news, article, or creative text generation
5. Information retrieval — building semantic search systems for Tatar
6. Linguistic research — analysing morphology, syntax, and lexicon of modern Tatar

---

▶️ Quick Start — Examples (Python)

Важно: перед запуском задайте HF-токен в переменной окружения Windows (PowerShell): setx HF_TOKEN "hf_...your_token..." Linux / macOS: export HF_TOKEN=hf_...your_token...

---

1) Loading Private Dataset (Standard Mode)

import os
from huggingface_hub import login
from datasets import load_dataset

# 1) Authentication (do this once per environment)
token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
login(token=token)

# 2) File paths
REPO = "arabovs-ai-lab/TatCorp_222M"

train_files = [
    f"hf://datasets/{REPO}@main/data/train-{i:05d}-of-00030.parquet"
    for i in range(30)
]
validation_files = [
    f"hf://datasets/{REPO}@main/data/validation-00000-of-00001.parquet"
]
test_files = [
    f"hf://datasets/{REPO}@main/data/test-00000-of-00001.parquet"
]

data_files = {
    "train": train_files,
    "validation": validation_files,
    "test": test_files,
}

# 3) Load dataset (auth handled automatically after login)
ds_dict = load_dataset("parquet", data_files=data_files)

# 4) Verification
for split, ds in ds_dict.items():
    print(f"{split}: {len(ds)} examples")

print("Columns:", ds_dict["train"].column_names)
print("First example:", ds_dict["train"][0])

---

2) Loading Private Dataset (Streaming)

import os
from huggingface_hub import login
from datasets import load_dataset

# --- Auth
token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
login(token=token)

# --- Streaming
REPO = "arabovs-ai-lab/TatCorp_222M"

train_files = [
    f"hf://datasets/{REPO}@main/data/train-{i:05d}-of-00030.parquet"
    for i in range(30)
]

ds_stream = load_dataset(
    "parquet",
    data_files={"train": train_files},
    split="train",
    streaming=True,
)

# --- Test
it = iter(ds_stream)
print(next(it))

---

3) Tokenisation for MLM (Example with XLM-R)

from transformers import (
    AutoTokenizer,
    AutoModelForMaskedLM,
    DataCollatorForLanguageModeling,
)

# assume ds_dict already loaded
ds = ds_dict["train"]

tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")

def tokenize_function(examples):
    return tokenizer(
        examples["content"],
        truncation=True,
        max_length=512,
    )

tokenized = ds.map(
    tokenize_function,
    batched=True,
    remove_columns=ds.column_names,
)

model = AutoModelForMaskedLM.from_pretrained("xlm-roberta-base")

data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=True,
    mlm_probability=0.15,
)

---

4) Preparing a Classifier (by Categories)

from transformers import AutoTokenizer, AutoModelForSequenceClassification

ds = ds_dict["train"]

labels = sorted(set(ds["category"]))
label2id = {label: i for i, label in enumerate(labels)}
id2label = {i: label for label, i in label2id.items()}

tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")

def preprocess(examples):
    tok = tokenizer(
        examples["title"],
        examples["content"],
        truncation=True,
        max_length=512,
    )
    tok["labels"] = [label2id[c] for c in examples["category"]]
    return tok

tokenized_clf = ds.map(
    preprocess,
    batched=True,
    remove_columns=ds.column_names,
)

model_clf = AutoModelForSequenceClassification.from_pretrained(
    "xlm-roberta-base",
    num_labels=len(labels),
    label2id=label2id,
    id2label=id2label,
)

---

5) Local Dataset Saving

from datasets import load_dataset

REPO = "arabovs-ai-lab/TatCorp_222M"

train_files = [
    f"hf://datasets/{REPO}@main/data/train-{i:05d}-of-00030.parquet"
    for i in range(30)
]

ds_local = load_dataset(
    "parquet",
    data_files={"train": train_files},
    split="train",
)

ds_local.save_to_disk("tatcorp_train_dataset")
print("Saved to tatcorp_train_dataset")

📚 References and Contacts

• Dataset page (private): https://huggingface.co/datasets/arabovs-ai-lab/TatCorp_222M
• Tokenisers: https://huggingface.co/arabovs-ai-lab/TatarTokenizers
• Embeddings: https://huggingface.co/arabovs-ai-lab/Tatar2Vec

Access requests / DUA / contact: cool.araby@gmail.com

🔬 About the Laboratory

Arabovs AI Research Lab — a private research laboratory of Associate Professor Mullosharaf Kurbonovich Arabov, specialising in Natural Language Processing, computational linguistics, and machine learning for low-resource languages.

Last updated: January 2025