domyn/Domyn-Edge-Base
Text Generation • 4B • Updated • 1.14k • 1
text stringlengths 41 4.3M | token_length int64 14 368k | src_path stringclasses 47
values |
|---|---|---|
### Code:
```python
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
import matplotlib.pyplot as plt # data visualization
import seaborn as sns
# Input data files are available in the read-only "../input/" directory
# For example, running this (by clickin... | 1,819 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
### Code:
```python
# **Comparison between different encoding methods**
# * LabelEncoder (sklearn)
# * OrdinalEncoder (sklearn)
# * One-hot Encoder (sklearn)
# * get_dummies (pandas)
# * map (pandas)
# The "titanic disaster" dataset is used for demonstration.
import numpy as np
import pandas as pd
import sklearn
pd... | 724 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# Students Performance in Exams
### Context
Marks secured by the students
### Content
This data set consists of the marks secured by the students in various subjects.
Kaggle dataset identifier: students-performance-in-exams
### Code:
```python
import numpy as np # linear algebra
import pandas as pd # data pro... | 1,669 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# Healthy Diet Breakdown
This is a small dataset that contains information about food groups along with serving information. The dataset contains 8 food items: Vegetables, Fruits, Grains, Dairy, meat, fish, nuts, and fats. For each food group, the following information is provided:
- Serving Per Day
- Serving Per We... | 9,883 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# 120 years of Olympic history: athletes and results
### Context
This is a historical dataset on the modern Olympic Games, including all the Games from Athens 1896 to Rio 2016. I scraped this data from www.sports-reference.com in May 2018. The R code I used to [scrape](https://github.com/rgriff23/Olympic_history/blo... | 3,643 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# Health care: Heart attack possibility
**About data set**
This database contains 76 attributes, but all published experiments refer to using a subset of 14 of them. In particular, the Cleveland database is the only one that has been used by ML researchers to
this date.The "target" field refers to the presence of hear... | 7,954 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
### Code:
```python
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
# Input data files are available in the read-only "../input/" directory
# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory... | 632 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# Sample Data for WordAndDoc2vec
Kaggle dataset identifier: sample-data-wordanddoc2vec
### Code:
```python
from IPython.display import SVG, Image
import os.path
import sys
import re
import itertools
import csv
import datetime
import pickle
import random
from collections import defaultdict, Counter
import gc
import ... | 6,137 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# NYC Open Data
### Context
NYC Open Data is an opportunity to engage New Yorkers in the information that is produced and used by City government. We believe that every New Yorker can benefit from Open Data, and Open Data can benefit from every New Yorker. Source: https://opendata.cityofnewyork.us/overview/
### Co... | 782 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
### Code:
```python
import pandas as pd
import numpy as np
from sklearn.linear_model import LogisticRegression
train_df = pd.read_csv("../input/titanic/train.csv")
test_df = pd.read_csv("../input/titanic/test.csv")
train_df.head()
test_df.head()
X = train_df["Pclass"].values.reshape(-1, 1)
y = train_df["Survived"].... | 243 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
### Code:
```python
import numpy as np
from keras import backend as K
from keras import models
from keras import layers
def make_data(batch_size, input_size, index):
"""
batch_size : Total batch size
input_size : Total input size, values are random number except at the attention_index
index : value... | 800 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# Predict CLTV of a customer
Problem Statement
VahanBima is one of the leading insurance companies in India. It provides motor vehicle insurances at best prices with 24/7 claim settlement. It offers different types of policies for both personal and commercial vehicles. It has established its brand across different... | 2,424 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# weight-height.csv
Kaggle dataset identifier: weight-height
### Code:
```python
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
# Input data files are available in the read-only "../input/" directory
# For example, running this (by clicking run or press... | 1,116 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
# Global Terrorism Database
# Context
Information on more than 180,000 Terrorist Attacks
The Global Terrorism Database (GTD) is an open-source database including information on terrorist attacks around the world from 1970 through 2017. The GTD includes systematic data on domestic as well as international terrorist ... | 11,145 | coding/issues_kaggle_notebooks/source=kaggle/9_ba9709b170b24067b526e441eeb8a169_000207_000000-0.parquet |
End of preview. Expand in Data Studio
Domyn Edge Pretraining Mix 8.4T
Phased release. Domyn Edge is a rolling release in three bundles. The first, released on October 20, 2026, includes this dataset,
domyn/Domyn-Edge-Baseanddomyn/Domyn-Edge-SFT. The data mixes for pre-training stages 2 to 6 follow in a later bundle. For an overview, see the launch blog post.
This is the stage 1 pre-training mix of Domyn Edge Base: 8.44 trillion tokens of web text, code, scientific papers, math and multilingual text. All counts refer to the tokens seen during training, measured with the Domyn Edge tokenizer.
This release contains the text of 6.95 trillion of those tokens. For the stack-edu code and for two NVIDIA datasets whose terms do not allow redistribution, it contains the upstream record identifier of each document instead of its text, as described in Documents published as identifiers.
Composition
The mix is drawn from five source families: web, code, STEM, multilingual and synthetic.
| Family | Source | Tokens | Share | In this release |
|---|---|---|---|---|
| Web | Web crawl | 3.65T | 43.20% | yes |
| Web | Wikipedia | 4.0B | 0.05% | yes |
| Synthetic | Synthetic | 1.68T | 19.91% | yes, Nemotron CC Math v1 and Nemotron Pretraining Code v2 as identifiers |
| STEM | Science PDFs | 1.16T | 13.74% | yes |
| STEM | Math | 130B | 1.54% | yes |
| STEM | Research papers | 78B | 0.93% | yes |
| Multilingual | Multilingual | 880B | 10.43% | yes |
| Code | Code | 860B | 10.20% | issues and Kaggle notebooks, stack-edu as identifiers |
Science PDFs and research papers were upsampled. The repeated copies are in the upsampling/
subfolders of data/science_pdfs/ and data/research_papers/, and they are included in the token
count.
Web crawl
| Source | Config | Tokens | Share of web crawl | Share of mix |
|---|---|---|---|---|
| Dolma CC | crawl_dolma-cc |
3.55T | 97.5% | 42.11% |
| Common Crawl, Domyn pipeline | crawl_fineweb |
92.4B | 2.5% | 1.10% |
Synthetic
| Source | Config | Tokens | Share of synthetic | Share of mix |
|---|---|---|---|---|
| Pre-training synthetic | synthetic_pretraining_synthetic |
1.05T | 62.8% | 12.50% |
| FinePhrase | synthetic_finephrase |
625.0B | 37.2% | 7.41% |
Nemotron Wiki-Rewrite (7.9B tokens of the pre-training synthetic data) is available in a separate
config, synthetic_wiki_rewrite; see Wikipedia attribution. Nemotron CC
Math v1 (119.1B tokens) and Nemotron Pretraining Code v2 (514.0B tokens) are also part of the
pre-training synthetic data, and they are published as identifiers (see
Omitted datasets). The table shows the full training mix.
Code
The issues and Kaggle notebooks are in the code_Issues-Kaggle config. The other languages are
stack-edu code, published as identifiers (see Stack-edu). The table shows the full
training mix.
| Source | Tokens | Share of code | Share of mix |
|---|---|---|---|
| Python | 218.8B | 25.4% | 2.59% |
| Java | 175.1B | 20.3% | 2.07% |
| C++ | 131.9B | 15.3% | 1.56% |
| Markdown | 109.2B | 12.7% | 1.29% |
| JavaScript | 56.6B | 6.6% | 0.67% |
| C# | 39.6B | 4.6% | 0.47% |
| C | 25.9B | 3.0% | 0.31% |
| PHP | 19.2B | 2.2% | 0.23% |
| Shell | 16.7B | 1.9% | 0.20% |
| TypeScript | 14.7B | 1.7% | 0.17% |
| SQL | 14.0B | 1.6% | 0.17% |
| Rust | 9.1B | 1.1% | 0.11% |
| Swift | 9.1B | 1.1% | 0.11% |
| Ruby | 8.5B | 1.0% | 0.10% |
Issues and Kaggle notebooks (code_Issues-Kaggle) |
7.7B | 0.9% | 0.09% |
| Go | 4.3B | 0.5% | 0.05% |
Python, Java and C++ are included in two forms: as published, and in our version that adds a
reasoning trace, in <think> tags, after the code. Our version accounts for 159.2B tokens of
Python, 123.8B of Java and 93.0B of C++. The other stack-edu languages are included as published.
Some languages were upsampled, and the repeated copies are included in the token count.
Multilingual
| Language | Tokens | Share of multilingual | Share of mix |
|---|---|---|---|
| Spanish | 116.4B | 13.2% | 1.38% |
| Italian | 115.9B | 13.2% | 1.37% |
| German | 114.9B | 13.1% | 1.36% |
| French | 112.7B | 12.8% | 1.34% |
| Portuguese | 92.1B | 10.5% | 1.09% |
| Russian | 63.0B | 7.2% | 0.75% |
| Chinese | 63.0B | 7.2% | 0.75% |
| Japanese | 42.0B | 4.8% | 0.50% |
| Turkish | 42.0B | 4.8% | 0.50% |
| Dutch | 42.0B | 4.8% | 0.50% |
| Indonesian | 42.0B | 4.8% | 0.50% |
| Arabic | 33.9B | 3.9% | 0.40% |
Each language has a high-quality subset (multilingual_<lang>_hq, from FineWeb2-HQ) and a
lower-quality subset (multilingual_<lang>_lowq, from HPLT 3.0), except Russian and Chinese, which
have only the high-quality subset. The FineWiki data for Indonesian, Japanese, Dutch, Portuguese and
Turkish is available in the wiki config, together with English Wikipedia; see
Wikipedia attribution.
Statistics
| Trained on | In this release | |
|---|---|---|
| Tokens | 8,438,903,495,207 | 6,953,078,932,001 |
| Documents | 5,881,410,200 | 4,426,534,737 |
How to use
Each source is available as a named config:
from datasets import load_dataset
italian = load_dataset(
"domyn/Domyn-Edge-Pretraining-Mix-8.4T",
"multilingual_ita_hq",
split="train",
streaming=True,
)
The default config loads the whole release except the Wikipedia data, which has three additional
columns and is available in two separate configs, wiki and synthetic_wiki_rewrite. To load the
full release, load default and both of these configs.
Data is stored under data/ as Parquet files of about 1 GB each, compressed with Snappy. Code,
crawl, synthetic and multilingual data are grouped in subfolders (data/code/, data/crawl/,
data/synthetic/, data/multilingual/); math, research_papers and science_pdfs are at the
top level of data/, and Wikipedia text is in data/wiki/ and data/synthetic/wiki_rewrite/. The
identifiers of the documents published without text are in data/synthetic-ids/ and
data/code-ids/, which are not part of any config (see
Documents published as identifiers).
Columns
| Column | Type | Description |
|---|---|---|
text |
string | The document. |
token_length |
int64 | Length of the document in Domyn Edge tokens. |
src_path |
string | The upstream corpus file the document came from, so any document can be traced back to its source. |
The Wikipedia configs have three additional columns, described in Wikipedia attribution.
Wikipedia attribution
Wikipedia text is licensed under CC BY-SA 4.0, which requires attribution to the contributors of
each article. The wiki and synthetic_wiki_rewrite configs include three additional columns that
identify the source of each document:
| Column | Type | Description |
|---|---|---|
source_id |
string | The upstream record: enwiki/<page id> for English Wikipedia, <wiki>/<page id> for FineWiki (for example ptwiki/25726), or the NVIDIA record uuid for Wiki-Rewrite. |
source_url |
string | The article URL. Null for Wiki-Rewrite. |
revision_id |
string | The article revision the text was taken from. Null for Wiki-Rewrite. |
To view a revision, open https://<language>.wikipedia.org/w/index.php?oldid=<revision_id>; the
contributors are listed in the article history. Wiki-Rewrite rows are attributed to the NVIDIA
record they come from, because NVIDIA does not link its records to the original articles. For
1,934 Wiki-Rewrite rows, all three columns are null, because the same text appears in more than one
NVIDIA record. Refer to ATTRIBUTION.md for details.
Sources
| Source | Link | Notes |
|---|---|---|
| Dolma 3 mix (6T) | allenai/dolma3_mix-6T | Most of the web crawl, the olmOCR science PDFs, the stack-edu code, FineMath, the RedPajama proof-pile arXiv papers and the Dolma 1.7 Wikipedia text. The stack-edu code is published as identifiers. |
| Common Crawl, Domyn pipeline | commoncrawl.org | Our FineWeb-style pipeline, run on CC-MAIN-2025-47. |
| FineWeb2-HQ | epfml/FineWeb2-HQ | High-quality multilingual subsets. |
| HPLT 3.0 | HPLT/HPLT3.0 | Lower-quality multilingual subsets. |
| FineWiki | HuggingFaceFW/finewiki | Five languages, in the wiki config. |
| FinePhrase | HuggingFaceFW/finephrase | |
| SYNTH | PleIAs/SYNTH | |
| Nemotron CC Math v1 | nvidia/Nemotron-CC-Math-v1 | The 4plus split, with reasoning traces we added. Published as identifiers. |
| Nemotron Pretraining Code v2 | nvidia/Nemotron-Pretraining-Code-v2 | Published as identifiers. |
| Nemotron Pretraining Specialized v1 | nvidia/Nemotron-Pretraining-Specialized-v1 | The Wiki-Rewrite subset is in the synthetic_wiki_rewrite config. |
| Nemotron Pretraining Specialized v1.1 | nvidia/Nemotron-Pretraining-Specialized-v1.1 | |
| OPC annealing corpus | OpenCoder-LLM/opc-annealing-corpus | |
| Swallow Code v2 | tokyotech-llm/swallow-code-v2 | |
| Issues and Kaggle notebooks | HuggingFaceTB/issues-kaggle-notebooks | The only code published as text. |
Documents published as identifiers
For two groups of documents, this release contains the identifier of the upstream record instead of the text. You can use these identifiers to rebuild the documents from the original datasets.
Omitted datasets
Two upstream datasets in the training mix are distributed under terms that do not allow redistribution, so this release does not contain their text. After you accept the upstream terms, you can rebuild their documents.
| Dataset | Revision | Subsets used | Tokens | Share of mix |
|---|---|---|---|---|
| Nemotron CC Math v1 | 397a2502 |
The 4plus split, with reasoning traces we added |
119.1B | 1.41% |
| Nemotron Pretraining Code v2 | 7b1a453d |
The synthetic code review, question answering, rewriting, student-teacher and transpilation subsets | 514.0B | 6.09% |
Stack-edu
The stack-edu code, 852.7B tokens or 10.10% of the mix, comes from the stack-edu subsets of the Dolma 3 mix. Source code requires per-file attribution, so for each file this release contains the identifier of the upstream record instead of the text. The records of the Dolma 3 mix include the repository, path and license of each file, which are required to attribute it.
| Dataset | Revision | Subsets used |
|---|---|---|
| Dolma 3 mix (6T) | 689a3ea2 |
data/stack_edu-<language> for C, C++, C#, Go, Java, JavaScript, Markdown, PHP, Python, Ruby, Rust, Shell, SQL, Swift and TypeScript |
Files and columns
The files are Parquet. The identifiers of the omitted datasets are under
data/synthetic-ids/pretraining_synthetic/<dataset>/, for example
data/synthetic-ids/pretraining_synthetic/Nemotron-CC-Math-v1/. The identifiers of the stack-edu
code are under data/code-ids/<language>/stack_edu/ for the files as published and
data/code-ids/<language>/stack_edu_traces/ for our version with reasoning traces. The repeated
copies of upsampled languages are in upsampling/ subfolders, for example
data/code-ids/Rust/upsampling/stack_edu/. To read the files, pass their folder to data_files:
code_ids = load_dataset(
"domyn/Domyn-Edge-Pretraining-Mix-8.4T",
data_files="data/code-ids/Rust/**/*.parquet",
split="train",
)
The files have the following columns:
| Column | Type | Description |
|---|---|---|
source_dataset |
string | The upstream dataset, for example nvidia/Nemotron-CC-Math-v1. |
source_subset |
string | The upstream folder that contains the record: 4plus, Synthetic-Code/synthetic-<subset> or data/stack_edu-<language>. |
source_id |
string | The upstream record: its id for Nemotron CC Math v1, its nemotron_pretraining_id for Nemotron Pretraining Code v2, and the Software Heritage blob id of the file for stack-edu. |
reasoning_trace |
string | Our reasoning trace, for Nemotron CC Math v1 documents and for the stack-edu files of our version with reasoning traces, up to 16k tokens. Null otherwise. |
token_length |
int64 | Length of the document as trained, in Domyn Edge tokens. |
src_path |
string | The upstream corpus file the document came from, as in the other configs. |
Each row is one document of the mix, so a record that was repeated in the mix appears once for each repetition. Documents with no text are not listed. In our version of the stack-edu code, files longer than 16k tokens have no trace.
Rebuilding the documents
To rebuild the documents:
- Accept the terms of each upstream dataset on its Hugging Face page. Nemotron Pretraining Code v2 requires approval by NVIDIA.
- Download the folders named in
source_subset, at the revisions listed in the tables above. - Keep one upstream record per identifier. The Dolma 3 mix repeats most stack-edu files up to five
times, with the same text. In the Dolma 3 mix, the blob id is the
blob_idfield of the Python files and theidfield of the files of the other languages. - Join the identifier files with the upstream records on
source_id. The upstream text is in thecontentcolumn for Nemotron Pretraining Code v2 and in thetextcolumn for the other datasets. - Where
reasoning_traceis not null, add it after the upstream text:text + "\n<think>" + reasoning_trace + "</think>".
The following example rebuilds the Nemotron CC Math v1 documents with polars:
import polars as pl
from huggingface_hub import snapshot_download
ids_dir = snapshot_download(
"domyn/Domyn-Edge-Pretraining-Mix-8.4T",
repo_type="dataset",
allow_patterns="data/synthetic-ids/pretraining_synthetic/Nemotron-CC-Math-v1/*",
)
math_dir = snapshot_download(
"nvidia/Nemotron-CC-Math-v1",
repo_type="dataset",
revision="397a2502f2028c659ba411a6c4935b464a7f03aa",
allow_patterns="4plus/*",
)
ids = pl.scan_parquet(f"{ids_dir}/data/synthetic-ids/pretraining_synthetic/Nemotron-CC-Math-v1/*.parquet")
upstream = (
pl.scan_parquet(f"{math_dir}/4plus/*.parquet")
.select(pl.col("id").alias("source_id"), "text")
.unique("source_id")
)
documents = ids.join(upstream, on="source_id").select(
(pl.col("text") + "\n<think>" + pl.col("reasoning_trace") + "</think>").alias("text"),
"token_length",
"src_path",
)
The following example rebuilds the Rust files:
import polars as pl
from huggingface_hub import snapshot_download
ids_dir = snapshot_download(
"domyn/Domyn-Edge-Pretraining-Mix-8.4T",
repo_type="dataset",
allow_patterns="data/code-ids/Rust/**",
)
dolma_dir = snapshot_download(
"allenai/dolma3_mix-6T",
repo_type="dataset",
revision="689a3ea2d8217e64d73a5058913fa43ad15e81aa",
allow_patterns="data/stack_edu-Rust/*",
)
ids = pl.scan_parquet(f"{ids_dir}/data/code-ids/Rust/**/*.parquet")
upstream = (
pl.scan_ndjson(f"{dolma_dir}/data/stack_edu-Rust/*.jsonl.zst")
.select(pl.col("id").alias("source_id"), "text")
.unique("source_id")
)
documents = ids.join(upstream, on="source_id").select("text", "token_length", "src_path")
Reasoning traces
The reasoning traces inside <think> tags were generated by us and are released under the Apache
License 2.0. They appear in the reasoning_trace column of the Nemotron CC Math v1 identifier files
and of the stack-edu identifier files for Python, Java and C++. The Apache 2.0 grant covers the
traces only, while the upstream documents each trace is attached to are covered by the license of
their respective source.
License
Our contributions to this dataset, including the curation and the reasoning traces, are released under the Apache License 2.0. Upstream corpora maintain their original license terms.
Read ATTRIBUTION.md before redistributing any part of this data. It lists every
upstream dataset with its publisher, repository and license. The LICENSES/ folder contains the
full text of each license: ODC-BY 1.0, CC BY 4.0, CC BY-SA 4.0, CC0 1.0, GFDL 1.3, Apache 2.0 and
the NVIDIA Data Agreement for Model Training.
Citation
The technical report is forthcoming in November 2026. Until it is published, please cite the dataset:
@misc{domyn_edge_pretraining_mix_2026,
title = {Domyn Edge Pretraining Mix},
author = {Domyn},
year = {2026},
note = {Technical report, forthcoming in November 2026.},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/domyn/Domyn-Edge-Pretraining-Mix-8.4T}
}
- Downloads last month
- 695