mlnomad/fineweb-edu-gemma4-1024

FineWeb-Edu — pre-tokenized for fast LM pretraining (Gemma tokenizer, ArrayRecord/Grain)

Pre-tokenized FineWeb-Edu (sample/100BT), packed into fixed-length sequences and stored as ArrayRecord shards for zero-overhead streaming with Grain. No on-the-fly tokenization at train time — you read int32 tokens straight off disk.

Format

• Tokenizer: google/gemma-4-12B-it (vocab size 262144). Documents are separated by the EOS token id 1.
• Packing: the token stream is concatenated (with EOS between documents) and split into fixed-length records. Each record is 1025 int32 tokens = 4100 bytes, i.e. seq_len = 1024 plus one extra token so you can build (input, target) by shifting.
• Layout: 90 shards, shard_00000.arrayrecord … shard_00089.arrayrecord, ~365M tokens each, **32B tokens total** (~128 GB).
• Each shard's records are independent and order within/across shards is fixed (deterministic), so training is reproducible and resumable by index.

Quick start (Grain)

import glob
import numpy as np
import grain
from huggingface_hub import snapshot_download

# 1. Download the shards locally (Grain reads local ArrayRecord files)
local = snapshot_download(
    "mlnomad/fineweb-edu-gemma4-1024",
    repo_type="dataset",
    allow_patterns=["*.arrayrecord"],
)
shards = sorted(glob.glob(f"{local}/*.arrayrecord"))

SEQ_LEN = 1024  # records are SEQ_LEN + 1 tokens

def decode(record_bytes):
    toks = np.frombuffer(record_bytes, dtype=np.int32)  # shape [1025]
    return {"input_ids": toks[:-1], "labels": toks[1:]}  # causal shift

dataset = (
    grain.MapDataset.source(grain.ArrayRecordDataSource(shards))
    .shuffle(seed=42)
    .map(decode)
    .batch(batch_size=32)
)

for batch in dataset:
    # batch["input_ids"], batch["labels"]: [32, 1024] int32
    ...

Streaming / resumable iteration

loader = grain.DataLoader(
    data_source=grain.ArrayRecordDataSource(shards),
    sampler=grain.IndexSampler(
        num_records=len(grain.ArrayRecordDataSource(shards)),
        shuffle=True, seed=42, num_epochs=None,
    ),
    operations=[grain.MapOperation(decode), grain.BatchOperation(32)],
    worker_count=16,  # parallel reader threads
)
it = iter(loader)
# loader.checkpoint()/restore lets you resume at the exact record index.

Reading a single shard without Grain

import numpy as np
from array_record.python.array_record_module import ArrayRecordReader

r = ArrayRecordReader("shard_00000.arrayrecord")
rec = r.read([0])[0]                       # bytes for record 0
toks = np.frombuffer(rec, dtype=np.int32)  # [1025] int32 token ids

Notes

• Built from the sample/100BT split of FineWeb-Edu. The trailing <1025-token remainder of each source file is dropped during packing (negligible).
• Token ids index the Gemma vocabulary; decode with the google/gemma-* tokenizer if you need text.

License & attribution

Derived from FineWeb-Edu (HuggingFaceFW/fineweb-edu), released under ODC-By 1.0. This tokenized derivative is provided under the same terms; please cite FineWeb-Edu.