Abzalbek89/corpus_clean_tokenized

Kazakh Tokenized Corpus (2048 blocks)

Pre-tokenized Kazakh corpus ready for language model training. Built from Abzalbek89/corpus_clean using Abzalbek89/kk-tokenizer-bpe-32k.

Dataset Summary

Metric	Value
Train blocks	236,981
Validation blocks	12,473
Total blocks	249,454
Block size	2,048 tokens
Total tokens	~0.51B (510M)
Tokenizer	ByteLevel BPE, 32K vocab
Val ratio	5%

Pipeline

1. Source: Abzalbek89/corpus_clean — 1,502,583 cleaned Kazakh texts
2. Chunking: Long texts split into chunks of up to 20K characters at paragraph/word boundaries (1,548,725 chunks)
3. Tokenization: ByteLevel BPE tokenizer (Abzalbek89/kk-tokenizer-bpe-32k, vocab=32,000)
4. Grouping: All token sequences concatenated and split into fixed blocks of 2,048 tokens (no padding, no truncation)
5. Split: 95% train / 5% validation (seed=42)

Fields

• input_ids (list[int]) — token IDs, length 2,048
• labels (list[int]) — copy of input_ids (for causal LM training)

Usage

from datasets import load_dataset

ds = load_dataset("Abzalbek89/corpus_clean_tokenized")

train = ds["train"]
val = ds["validation"]

# Each example is a fixed-size block of 2048 tokens
print(len(train[0]["input_ids"]))   # 2048
print(len(train[0]["labels"]))      # 2048

Training with HuggingFace Trainer

from transformers import AutoModelForCausalLM, TrainingArguments, Trainer
from datasets import load_dataset

ds = load_dataset("Abzalbek89/corpus_clean_tokenized")

model = AutoModelForCausalLM.from_pretrained("your-model")

training_args = TrainingArguments(
    output_dir="./output",
    per_device_train_batch_size=8,
    num_train_epochs=3,
    learning_rate=5e-4,
    logging_steps=100,
    save_steps=1000,
    eval_strategy="steps",
    eval_steps=500,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=ds["train"],
    eval_dataset=ds["validation"],
)

trainer.train()

Source Data

Component	Link
Raw corpus	Abzalbek89/corpus_clean
Tokenizer	Abzalbek89/kk-tokenizer-bpe-32k

License

Apache 2.0