@inproceedings{zhang-etal-2026-huggingbrew,
title = "{H}ugging{B}rew: Brewing High-Quality Mid-Training Data from the {H}ugging{F}ace Community Datasets",
author = "Zhang, Han and
Chen, Ningqi and
Liu, Xiao and
Lin, Zhenghao and
Gong, Yeyun and
Lu, Yan and
Yang, Hanfang",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1519/",
pages = "32346--32365",
ISBN = "979-8-89176-462-0",
abstract = "Data curation for large language models (LLMs) has been extensively studied for both the pre-training and post-training stages, yet the intermediate $\textit{mid-training}$ phase{---}bridging general pretraining and task-specific fine-tuning{---}remains critically underexplored. Most existing mid-training data for state-of-the-art models remains proprietary. Meanwhile, the Hugging Face platform hosts a huge amount of community-contributed datasets, yet no systematic effort has been made to refine this rich ecosystem into a unified high-quality corpus. We present $\textbf{HuggingBrew}$, a large-scale mid-training corpus constructed through a principled, multi-step curation pipeline applied to the massive publicly accessible Hugging Face datasets. The pipeline operates at two complementary granularities: $\textit{dataset-level}$ operations act on each dataset independently, while $\textit{corpus-level}$ operations are applied jointly across all retained corpus. We construct three progressively curated variants{---}$\textsf{Draught}$, $\textsf{Craft}$, and $\textsf{Reserve}${---}containing 1.6T, 450B, and 40B tokens, respectively. Continual pretraining experiments under a 100B-token budget show that series yields substantial improvements over baselines across mathematical reasoning, general knowledge, and code generation, validating the untapped value of community-contributed data for mid-training."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhang-etal-2026-huggingbrew">
<titleInfo>
<title>HuggingBrew: Brewing High-Quality Mid-Training Data from the HuggingFace Community Datasets</title>
</titleInfo>
<name type="personal">
<namePart type="given">Han</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ningqi</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiao</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhenghao</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yeyun</namePart>
<namePart type="family">Gong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yan</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanfang</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Data curation for large language models (LLMs) has been extensively studied for both the pre-training and post-training stages, yet the intermediate mid-training phase—bridging general pretraining and task-specific fine-tuning—remains critically underexplored. Most existing mid-training data for state-of-the-art models remains proprietary. Meanwhile, the Hugging Face platform hosts a huge amount of community-contributed datasets, yet no systematic effort has been made to refine this rich ecosystem into a unified high-quality corpus. We present HuggingBrew, a large-scale mid-training corpus constructed through a principled, multi-step curation pipeline applied to the massive publicly accessible Hugging Face datasets. The pipeline operates at two complementary granularities: dataset-level operations act on each dataset independently, while corpus-level operations are applied jointly across all retained corpus. We construct three progressively curated variants—Draught, Craft, and Reserve—containing 1.6T, 450B, and 40B tokens, respectively. Continual pretraining experiments under a 100B-token budget show that series yields substantial improvements over baselines across mathematical reasoning, general knowledge, and code generation, validating the untapped value of community-contributed data for mid-training.</abstract>
<identifier type="citekey">zhang-etal-2026-huggingbrew</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1519/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>32346</start>
<end>32365</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T HuggingBrew: Brewing High-Quality Mid-Training Data from the HuggingFace Community Datasets
%A Zhang, Han
%A Chen, Ningqi
%A Liu, Xiao
%A Lin, Zhenghao
%A Gong, Yeyun
%A Lu, Yan
%A Yang, Hanfang
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F zhang-etal-2026-huggingbrew
%X Data curation for large language models (LLMs) has been extensively studied for both the pre-training and post-training stages, yet the intermediate mid-training phase—bridging general pretraining and task-specific fine-tuning—remains critically underexplored. Most existing mid-training data for state-of-the-art models remains proprietary. Meanwhile, the Hugging Face platform hosts a huge amount of community-contributed datasets, yet no systematic effort has been made to refine this rich ecosystem into a unified high-quality corpus. We present HuggingBrew, a large-scale mid-training corpus constructed through a principled, multi-step curation pipeline applied to the massive publicly accessible Hugging Face datasets. The pipeline operates at two complementary granularities: dataset-level operations act on each dataset independently, while corpus-level operations are applied jointly across all retained corpus. We construct three progressively curated variants—Draught, Craft, and Reserve—containing 1.6T, 450B, and 40B tokens, respectively. Continual pretraining experiments under a 100B-token budget show that series yields substantial improvements over baselines across mathematical reasoning, general knowledge, and code generation, validating the untapped value of community-contributed data for mid-training.
%U https://aclanthology.org/2026.findings-emnlp.1519/
%P 32346-32365
Markdown (Informal)
[HuggingBrew: Brewing High-Quality Mid-Training Data from the HuggingFace Community Datasets](https://aclanthology.org/2026.findings-emnlp.1519/) (Zhang et al., Findings 2026)
ACL