---
license: odc-by
language:
  - en
task_categories:
  - text-generation
size_categories:
  - 1B<n<10B
tags:
  - scientific
  - pretraining
  - web-corpus
  - stem
  - common-crawl
---

# HainaWeb-Sci Sample (Anonymous Submission)

This repository hosts a **representative sample** of the HainaWeb-Sci scientific
web corpus, released anonymously for double-blind peer review at NeurIPS 2026.

The full 1.09T-token corpus, finalised license, classifier checkpoints, and
long-term maintenance repository will be released at the non-anonymous
location at camera-ready time, pending internal release review.

## Sample Composition

The sample (~3.46 GB compressed) is uniformly distributed across all
**14 STEM disciplines** covered by HainaWeb-Sci:

AerospaceEngineering, Agriculture, Astronomy, Biology, Chemistry,
CivilEngineering, ComputerScience, EarthScience,
ElectricalElectronicEngineering, MaterialsScience, Mathematics,
MechanicalEngineering, Medicine, Physics.

Each discipline lives under its own top-level folder; documents are stored
as gzip-compressed JSONL shards (`HainaWeb-Sci_<discipline>_<id>.jsonl.gz`).

## Data Schema

Each line is a JSON object. Refer to the paper's Appendix A (Datasheet) for
the full field-level specification. Key fields:

- `text`: main document text after the full curation pipeline.
- `metadata`: original WARC fields (URI, content-type, crawl date, etc.).
- `websci_meta.language`: detected language and confidence.
- `websci_meta.discipline`: fine-grained discipline distribution from
  the two-tier discipline router (top-4 disciplines with confidence).
- `websci_meta.model_quality_score`: content-driven fastText quality score.
- `websci_meta.sci_quality_score`: 0–5 scientific value score from the
  rubric-based quality classifier.

## Data Curation Pipeline
The HainaWeb-Sci corpus is constructed from DCLM-Pool and DCLM-Pool-aligned Common Crawl data (2013–March 2026). We process the raw data through a four-stage data-centric curation pipeline tailored to scientific web data:

- Stage 1 (Data Preparation) transforms raw web data into a consistent textual foundation through WARC parsing, text extraction, URL filtering, and language identification. 
- Stage 2 (Quality Filtering) focuses on science-aware content refinement by combining rule-based and model-based filtering. Unlike general-purpose pipelines, both heuristic rules and learned quality models are customized to better preserve scientifically meaningful content, particularly symbolic expressions and structured reasoning.
- Stage 3 (Deduplication) removes redundant content at scale to improve corpus diversity and training efficiency. 
- Stage 4 (Discipline-Specific Curation) introduces fine-grained discipline classification and scientific value scoring, enabling structured organization of scientific knowledge across 14 disciplines and providing quantitative signals for data selection.

## PII Handling

Personally identifiable information has been masked using the standard
Datatrove PII formatter: email addresses are replaced with placeholders
(`user@domain.org`); public IPs are replaced with the RFC 5737 TEST-NET
address (`198.51.100.1`).

## License

This sample is released under the
**Open Data Commons Attribution License (ODC-By) v1.0**.
Use of the data is also subject to the original
[Common Crawl terms of use](https://commoncrawl.org/terms-of-use).

## Intended Use

This anonymous sample is provided **solely for reviewer evaluation** of the
NeurIPS 2026 submission. Redistribution prior to camera-ready is not
permitted. The full corpus, code, classifier checkpoints, and Datasheet
will be released at the non-anonymous repository at camera-ready.

## Limitations

- English-only; humanities and social sciences are out of scope.
- Code and multimodal content are not included.
- Sample size is intentionally limited for reviewer inspection; quality
  metrics reported in the paper are computed on the full corpus, not
  this sample.

## Reporting Issues

Anonymous review channel — please contact the authors via the
OpenReview submission discussion thread.