metadata
license: odc-by
language:
- en
The Dump
The Dump is a diverse, general-purpose collection of English-language text datasets released under permissive licenses. Its primary purpose is to aggregate a wide variety of content types to support flexible language model training and experimentation.
Dataset Composition
To ensure a balanced sample, 110 000 rows were extracted from each source dataset using the configuration below:
| Dataset | Subset | Split | Column |
|---|---|---|---|
HuggingFaceFW/fineweb |
default | train | text |
HuggingFaceFW/fineweb-edu |
default | train | text |
allenai/c4 |
en | train | text |
agentlans/common-crawl-sample |
en | train | text |
agentlans/LLM360-TxT360-sample |
eng_Latn | train | text |
HuggingFaceFW/finepdfs |
eng_Latn | train | text |
agentlans/PleIAs-common_corpus-sample |
English | train | text |
EleutherAI/the_pile_deduplicated |
none | train | text |
DKYoon/SlimPajama-6B |
none | train | text |
tiiuae/falcon-refinedweb |
none | train | content |
agentlans/dolma-1m |
none | train | text |
agentlans/allenai-dolma3_mix-6T-sample |
none | train | text |
agentlans/BAAI-IndustryCorpus2-sample |
none | train | text |
agentlans/m-a-p-FineFineWeb-sample |
none | train | text |
Preprocessing & Cleaning
The data underwent the following preprocessing steps:
- PII Removal: Stripping of common personally identifiable information (PII) such as email addresses and URLs.
- Normalization: Unicode normalization applied to NFC format.
- Whitespace Cleanup: Leading and trailing spaces stripped.
- Deduplication: Texts were deduplicated via exact match.
Data Structure
The dataset is provided in JSONLines format, with each record structured as follows:
id: Unique UUID for the entry.text: The raw text content.source: The originating Hugging Face dataset name.
Limitations & Considerations
- Variable Quality & Formatting: Because these datasets originate from open web crawls, formatting, length, and text quality vary significantly.
- Unfiltered Content: As is typical with web text collections, the corpus may occasionally contain offensive or biased language. Users are advised to perform downstream filtering based on their specific use case.
Licence
Released under the ODC-BY (Open Data Commons Attribution) licence. If you have specific compliance requirements, please consult the individual source datasets linked above.