br-llm-data/Carolina_Wikipedia26TF_HIGH_PLUS_2M

Carolina + Wikipedia Rephrases — 2M

This dataset contains every record of the source HIGH dataset and a nested selection of previously absent Wikipedia reformulations.

Provenance

• Source HIGH: https://huggingface.co/datasets/br-llm-data/Carolina_Wikipedia26TF_HIGH
• Source rephrases: https://huggingface.co/datasets/br-llm-data/wikipedia_rephrase_pt
• Newly introduced domain: wikipedia_rephrase_remaining

Documents and actual tokenizer tokens

Tokenizer: Qwen/Qwen3-4B; no truncation, padding or special tokens.

Split	Documents before	Documents added	Documents after	Tokens before	Tokens added	Tokens after
train	5,908,040	1,900,000	7,808,040	5,330,069,550	952,911,500	6,282,981,050
validation	310,784	100,000	410,784	280,314,499	50,102,435	330,416,934
TOTAL	6,218,824	2,000,000	8,218,824	5,610,384,049	1,003,013,935	6,613,397,984

Sampling and split policy

• SHA-256 deduplication using the exact full text against HIGH and other rephrases.
• Only the nine reformulation subsets (not original) were eligible.
• Seeded random sampling proportional to each eligible subset's available records.
• Versions are nested: 1M ⊂ 2M ⊂ ALL (including train/validation membership).
• Original HIGH records preserve their split membership, but are re-ordered for domain mixing.
• Added examples are allocated ~95/5 across train/validation within each subset.
• Every batch approximates the NEW global stratum proportions.
• Batch ordering is physical in the dataset; dataloader reshuffling removes this order.
• See construction_report.json for counts per source subset and auditing.

Licensing

This is a composite dataset. Retain and check each component's original license, including Wikipedia-derived content attribution / share-alike requirements. Do not assume a single license for Carolina and Wikipedia.