lablup/dflash-training-enkocode-20b

DFlash Training EnKoCode 20B

Raw text training dataset for DFlash experiments, built from local Nemotron and Solar-derived corpora. The dataset is intentionally stored as untokenized Parquet so training code can load it with datasets.load_dataset, apply the target model tokenizer or chat template at training time, and avoid baking in one tokenizer's assumptions.

from datasets import load_dataset

ds = load_dataset("lablup/dflash-training-enkocode-20b")

Splits

split	records	characters
train	31,225,272	80,000,000,611
validation	39,152	80,070,910
total	31,264,424	80,080,071,521

The train split is approximately 80B raw characters, not exactly 20B tokens. Token counts will depend on the tokenizer and chat template used during training.

Composition

split	source_group	records	characters
train	nemotron_general_instruction	1,715,121	20,000,000,129
train	nemotron_science_math	374,630	8,000,000,408
train	nemotron_code_swe	245,194	6,000,030,865
train	solar_english	793,772	6,000,000,540
train	solar_ko_general	11,148,355	20,000,002,859
train	solar_ko_synthetic	12,647,133	12,000,000,328
train	solar_ko_domain	4,301,067	8,000,005,482
validation	nemotron_general_instruction	12,268	20,001,179
validation	nemotron_science_math	296	8,017,300
validation	nemotron_code_swe	242	6,006,713
validation	solar_english	1,192	6,003,780
validation	solar_ko_general	8,153	20,000,751
validation	solar_ko_synthetic	12,790	12,000,064
validation	solar_ko_domain	4,211	8,001,123

There is no dedicated Solar Korean code slice in the local source inventory used for this build. Korean code content, if present, is incidental inside the Korean general, synthetic, or domain sources.

Schema

Each row contains:

• id: stable exported row id
• text: raw text payload
• split: train or validation
• language: coarse language label, usually en or ko
• source: source dataset family
• source_group: composition bucket used during export
• category: optional category inferred from the source
• source_id: optional upstream row identifier
• source_path: relative source file path
• line_no: line number for JSONL-backed sources, if available
• row_idx: row index for table-backed sources, if available
• name: optional source name
• char_count: character count of text

Intended Use

This dataset is meant to be tokenized inside downstream training jobs:

from datasets import load_dataset
from transformers import AutoTokenizer

ds = load_dataset("lablup/dflash-training-enkocode-20b", split="train")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")

def tokenize(example):
    return tokenizer(example["text"], truncation=True)

tokenized = ds.map(tokenize, remove_columns=ds.column_names)

For chat-style models, format rows into the model's message schema and call tokenizer.apply_chat_template(...) during preprocessing.