JonathanMiddleton/fineweb-edu-dedup-shuffled-pretokenized

FineWeb-Edu-Dedup Shuffled Pretokenized

Pretokenized training shards built from a globally shuffled version of FineWeb-Edu-Dedup. Ready for direct consumption by the Daisy pretraining loop.

Summary

Property	Value
Total tokens	181,465,257,766 (~181.5B)
Train tokens	~180.5B
Val tokens	1,000,000,000 (1B)
Train shards	1,994
Val shards	10
Tokens per shard	100,000,000 (full shards); last shard per worker may be partial
Documents (train)	180,185,493
Documents (val)	992,927
Tokenizer	jonathanmiddleton/daisy (49,152 vocab, BPE)
Token dtype	uint16
Shard format	v3 (magic=20260114, version=3)
EOS token ID	49131

Directory Structure

train/
  000000.bin
  000001.bin
  ...
  001993.bin
val/
  000000.bin
  000001.bin
  ...
  000009.bin

Each .bin file contains a 1024-byte header followed by a flat array of uint16 token IDs.

Shard Format

Each shard file has a fixed 1024-byte header (256 int32 words) followed by the token payload:

Header word	Field	Value
0	magic	20260114
1	version	3
2	num_tokens	number of tokens in this shard
3	tokenizer_crc	CRC32 of tokenizer name (stored as uint32 in int32 slot)
4	vocab_size	49152
5	eos_id	49131
6	dtype_bits	16

The token stream is a concatenation of documents separated by EOS tokens:

[EOS] [doc1_token1] [doc1_token2] ... [EOS] [doc2_token1] ...

Every document begins with an EOS token (ID 49131). Documents may span shard boundaries: a document that doesn't fit entirely in one shard continues at the start of the next shard within the same worker's output. The training data loader treats all shards as a single continuous token stream.

Provenance

This dataset was produced by a two-stage pipeline:

Stage 1: Global Shuffle (parquet)

The 190,168,005 rows of HuggingFaceTB/smollm-corpus (fineweb-edu-dedup subset) were globally shuffled using a Fisher-Yates permutation with BLAKE2b-seeded PCG64 PRNG (seed=42). The shuffled parquet is published separately at JonathanMiddleton/fineweb-edu-dedup-shuffled.

The shuffle eliminates temporal and topical clustering from the upstream Common Crawl dump ordering, improving gradient diversity during pretraining.

Stage 2: Pretokenization (this dataset)

The shuffled parquet was tokenized using the jonathanmiddleton/daisy tokenizer (49,152 vocab BPE) and written as uint16 binary shards.

• Train/val split: The first 20 of 381 shuffled parquet files (5%) were reserved for validation. The remaining 361 files were used for training.
• Train shards: 190 parallel workers drained all 361 train parquet files, producing 1,994 shards (1,803 full shards of 100M tokens + 191 partial final shards).
• Val shards: 1 worker tokenized the 20 val parquet files, capped at 10 shards (1B tokens). Not all val documents were tokenized due to the shard cap.

Verification

Post-build validation confirmed:

• All shard headers are valid (magic, version, tokenizer CRC, payload size).
• Sequential shard naming with no gaps.
• Train EOS token count (180,185,493) matches the source row count for the 361 train parquet files (180,185,425 rows). The +68 difference is within tolerance (0.00004%), likely from documents whose tokenized content incidentally contains the EOS token ID.

Usage

Download

python -m data.download_dataset fineweb-edu-shuffled

This downloads to data/fineweb-edu-shuffled/train/ and data/fineweb-edu-shuffled/val/.

Training Configuration

In a Daisy training YAML config:

train_shards:
  - type: "fineweb_edu_shuffled"
    path: "data/fineweb-edu-shuffled/train"
    sequence_length: 65536

val_shards:
  - type: "fineweb_edu_shuffled"
    path: "data/fineweb-edu-shuffled/val"
    target_tokens: 1_000_000
    sequence_length: 65536

The data loader globs *.bin from the directory and reads shards sequentially.

Shard Range Selection

To use a subset of shards (e.g., for multi-stage training that avoids data reuse):

path: "data/fineweb-edu-shuffled/train[000500:001000]"

This selects shards 000500.bin through 001000.bin (inclusive), using the range filter supported by the Daisy data loader.

License

This dataset inherits the ODC-BY 1.0 license from FineWeb via SmolLM-Corpus.