MaagDeveloper/neo-foundation-mix-v1

Neo Foundation Mix v1

A multi-domain, raw-text corpus for pretraining small and experimental language models from scratch.

Neo Foundation Mix v1 is a curated mixture of educational web text, synthetic educational content, Python code, mathematics, general web text, and short stories.

The dataset was assembled to support general-purpose causal language model pretraining, with an emphasis on diverse knowledge and language patterns rather than conversational instruction tuning.

Dataset Overview

Property	Value
Dataset type	Raw text
Primary language	English
Training objective	Causal Language Modeling (Next Token Prediction)
Total documents	1,176,399
Training documents	1,161,634
Validation documents	3,000
Test documents	11,765
Dataset size	Approximately 4.17 GB (uncompressed)
Tokenizer	Not included
Token count	Not measured
Document separator	<|endoftext|> for optional text export

The dataset is tokenizer-independent and does not require any specific model architecture.

Data Sources

The corpus combines six sources:

Source	Purpose	Target proportion by character count
FineWeb-Edu	General knowledge and educational web content	55%
Cosmopedia v2	Synthetic educational text, explanations, and stories	15%
Python code	Programming and source-code patterns	10%
FineMath	Mathematics and mathematical explanations	10%
OpenWebText	General web text and linguistic diversity	5%
TinyStories	Simple language, narrative structure, and coherence	5%

These proportions represent the intended composition during dataset preparation. Exact token-level proportions depend on the tokenizer used.

Original Dataset References

• SmolLM-Corpus
• FineMath
• OpenWebText
• TinyStories
• CodeParrot Clean

Dataset Structure

Each document contains the following fields:

Field	Description
text	Raw document text
source	Source category
doc_id	Document identifier assigned during preparation
char_count	Number of characters
license	License-related metadata inherited or recorded during collection
origin_id	Original source identifier, where available

The dataset provides three predefined splits: train, validation, and test.

Preparation Pipeline

The preparation process includes:

1. Streaming collection from the source datasets.
2. Basic text normalization and filtering.
3. Document-level exact deduplication.
4. Character-budget-based sampling across sources.
5. Train, validation, and test splitting.
6. Export to Hugging Face Dataset format.

Long documents may have been truncated to respect sampling limits.

Exact deduplication does not eliminate all near-duplicates, paraphrases, or overlapping source content.

Usage

Load the complete dataset:

from datasets import load_dataset

dataset = load_dataset("MaagDeveloper/neo-foundation-mix-v1")

print(dataset)

Load only the training split:

train = load_dataset(
    "MaagDeveloper/neo-foundation-mix-v1",
    split="train"
)

Export to Plain Text

from datasets import load_dataset

train = load_dataset(
    "MaagDeveloper/neo-foundation-mix-v1",
    split="train"
)

with open("input.txt", "w", encoding="utf-8") as f:
    for row in train:
        f.write(row["text"].strip())
        f.write("<|endoftext|>\n")

The <|endoftext|> marker is used as an optional document separator in exported text files. It is not required to be a special token in the training tokenizer.

Intended Uses

This dataset is intended for:

• Pretraining decoder-only language models from scratch.
• Small language model research.
• Tokenizer training and evaluation.
• Training-data mixture experiments.
• Continued language-model pretraining.
• Research on language modeling and internal representations.

It is not specifically optimized for instruction following. Supervised instruction tuning may be performed separately.

Limitations

• The dataset primarily contains English-language content.
• Some source material is synthetically generated.
• Web content may contain factual errors, biases, or outdated information.
• Source documents may overlap semantically despite exact deduplication.
• Some documents may have been truncated.
• Token counts depend on the tokenizer and have not been independently measured.
• General-purpose training data does not guarantee strong downstream performance.

Licensing and Attribution

This dataset combines material originating from several independently maintained datasets.

Original sources may have different licenses, attribution requirements, and redistribution conditions.

The license column provides source-level metadata where available, but it should not be treated as a complete legal determination.

Users should consult the original dataset repositories and applicable terms before redistributing or commercially using the content.

Maintainer

MaagDeveloper

Version: v1

This dataset is intended as a reusable foundation for experiments with small language models and general-purpose pretraining.