metadata

language:
  - en
license: mit
task_categories:
  - text-generation
  - fill-mask
  - text-classification
  - retrieval
tags:
  - pretraining
  - language-modeling
  - encoder
  - decoder
  - foundation-model
  - transformer

Ettin Pre-training Data

Phase 1 of 3: Diverse pre-training data mixture (1.7T tokens) used to train the Ettin model suite.

This dataset contains the pre-training phase data used to train all Ettin encoder and decoder models. The data is provided in MDS format ready for use with Composer and the ModernBERT training repository.

📊 Data Composition

Data Source	Tokens (B)	Percentage	Description
DCLM	837.2	49.1%	High-quality web crawl data
CC Head	356.6	20.9%	Common Crawl head documents
Starcoder	263.9	15.5%	Code repositories and files
Reddit	80.3	4.7%	Social discussion threads
PeS2o	57.3	3.4%	Scientific papers
Arxiv	28.0	1.6%	Academic preprints
StackExchange	19.6	1.2%	Q&A forums
Tulu Flan	16.6	1.0%	Instruction-following data
Open-Web-Math	12.7	0.7%	Mathematical content
Algebraic StackExchange	12.6	0.7%	Math Q&A
CC News	7.3	0.4%	News articles
Wikipedia	7.3	0.4%	Encyclopedia articles
Total	1,704.7	100.0%	Diverse mixture for foundation training

🚀 Usage

For pre-training, see the ModernBERT repo: https://github.com/AnswerDotAI/ModernBERT

Direct Access

from streaming import StreamingDataset

# Load the streaming dataset
dataset = StreamingDataset(
    remote='https://huggingface.co/datasets/jhu-clsp/ettin-pretraining-data',
    local='/tmp/ettin-pretraining-data',
    shuffle=True
)

# Access samples
for sample in dataset:
    text = sample['text']
    # Process your data...

📁 Structure

Each folder contains one data source in MDS (Mosaic Data Shard) format:

arxiv/ - Academic papers from ArXiv
books/ - Literature and reference books
cc_head/ - High-quality Common Crawl documents
cc_news/ - News articles from Common Crawl
dclm/ - DataComp-LM filtered web data
open_web_math/ - Mathematical web content
algebraic_stackexchange/ - Math Q&A from StackExchange
pes2o/ - Scientific papers (PeS2o dataset)
reddit/ - Reddit discussion threads
stackexchange/ - General StackExchange Q&A
starcoder/ - Code from GitHub repositories
tulu_flan/ - Instruction-following examples
wikipedia/ - Wikipedia articles

🔗 Related Resources

Models: Ettin Model Suite (17M-1B parameters)
Phase 2: Mid-training Data (250B tokens)
Phase 3: Decay Phase Data (50B tokens)
Training Order: Batch-level Data Order
Paper: Arxiv link
Code: GitHub Repository

Citation

@misc{weller2025seqvsseqopen,
      title={Seq vs Seq: An Open Suite of Paired Encoders and Decoders}, 
      author={Orion Weller and Kathryn Ricci and Marc Marone and Antoine Chaffin and Dawn Lawrie and Benjamin Van Durme},
      year={2025},
      eprint={2507.11412},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.11412}, 
}