juliannunezb/mixed-pretrain-10b-gpt2

Mixed Pretraining 10B (GPT-2 BPE)

A 10-billion-token pretraining dataset, GPT-2 BPE tokenized, assembled as a diverse mix of web text, books, Wikipedia, code, academic papers, Q&A and instruction-formatted conversations.

Built to train a ~500M parameter from-scratch GPT-2-style transformer (see juliannunezb/transformer-lm-500m).

Mix

Source	Mix %	Tokens	Notes
fineweb	40.4%	4,039,999,700	reused from kjj0/fineweb10B-gpt2
fineweb_edu	15.2%	1,514,999,900	reused from karpathy/fineweb-edu-100B-gpt2-token-shards
openwebtext	10.1%	1,009,999,900	reused from chanind/openwebtext-gpt2
pg19	10.1%	1,009,999,900	tokenized from manu/project_gutenberg (en) — Project Gutenberg books
wikipedia	7.6%	757,499,900	tokenized from wikimedia/wikipedia config 20231101.en
github	7.6%	757,499,900	tokenized from codeparrot/github-code-clean (Python/JavaScript/Go)
stackexchange	5.1%	505,000,000	tokenized from HuggingFaceH4/stack-exchange-preferences (Q + top-scoring A)
instruct	4.0%	405,000,000	tokenized from teknium/OpenHermes-2.5 + databricks/databricks-dolly-15k
Total	100.0%	9,999,999,200

Format

Each .bin file is an llm.c-compatible shard:

[ header: 256 × int32 ]
  offset  0: magic    = 20240520
  offset  4: version  = 1
  offset  8: n_tokens = N
  offset 12..1023: reserved (zeros)
[ body: N × uint16  token IDs, each < 50257 ]

• Tokenizer: GPT-2 BPE via tiktoken.get_encoding("gpt2") (vocab size 50257).
• <|endoftext|> (id 50256) separates documents within a shard and also marks the boundary between sources inside a mixed shard.
• Each shard is exactly 100,000,000 tokens (100M), except possibly the last.
• Total shards: 100 → 10,000,000,000 tokens.

Loading

import numpy as np
from huggingface_hub import hf_hub_download

def load_shard(shard_idx: int) -> np.ndarray:
    path = hf_hub_download(
        repo_id="juliannunezb/mixed-pretrain-10b-gpt2",
        filename=f"mix_{shard_idx:06d}.bin",
        repo_type="dataset",
    )
    with open(path, "rb") as f:
        header = np.frombuffer(f.read(1024), dtype=np.int32)
        assert header[0] == 20240520 and header[1] == 1
        n = int(header[2])
        tokens = np.frombuffer(f.read(n * 2), dtype=np.uint16)
    return tokens

shard = load_shard(0)     # first shard: ~100M uint16 token IDs
print(shard[:20])         # e.g. array([ 464,  368, 2321, ...], dtype=uint16)

Decode back to text with tiktoken:

import tiktoken
enc = tiktoken.get_encoding("gpt2")
print(enc.decode(shard[:200].tolist()))

Reproducibility

Built with tokenize_and_mix.py on 2026-04-21 20:24:01 UTC. Tokenization ran against the HEAD of each source at that time.

• Pre-tokenized shards copied (no re-encoding) from:
  • kjj0/fineweb10B-gpt2
  • karpathy/fineweb-edu-100B-gpt2-token-shards
  • chanind/openwebtext-gpt2
• Fresh tokenized from:
  • manu/project_gutenberg (split en) — Project Gutenberg books
  • wikimedia/wikipedia (config 20231101.en)
  • codeparrot/github-code-clean (configs Python-all, JavaScript-all, Go-all)
  • HuggingFaceH4/stack-exchange-preferences — Q + top-scoring answer text
  • teknium/OpenHermes-2.5
  • databricks/databricks-dolly-15k

Formatting notes

• Documents within each source are separated by <|endoftext|> (50256).
• Code documents are prefixed with # Path: <repo_path>\n to provide structural context and are restricted to .py, .js, .ts, .go.
• Instruction documents follow a simple chat template:

  User: <instruction>
  Assistant: <response>
  

  Dolly entries that include a context field embed it between the user turn and the assistant turn (see the pipeline script for exact formatting).
• Within each mixed shard, sources appear in the fixed order: fineweb → fineweb_edu → openwebtext → pg19 → wikipedia → github → arxiv → stackexchange → instruct, separated by <|endoftext|> at each boundary.

Sanity checks run before upload

• 0 ≤ token_id < 50257 on every shard (asserts in writer).
• EOT counts sampled on first/middle/last shard — see sanity.json.
• Random 500-token decodes from 3 shards — see sanity.json for excerpts.

License

This dataset is a derivative of multiple sources. The most restrictive upstream license governs each subset:

Source	License
FineWeb / FineWeb-Edu	ODC-By 1.0
OpenWebText	CC0
PG-19	Apache 2.0 (pre-1919 public-domain books)
Wikipedia	CC-BY-SA 3.0 / 4.0
RedPajama github	permissive licenses (MIT/BSD/Apache) upstream
RedPajama arxiv	arXiv terms of use
RedPajama stackexchange	CC-BY-SA 4.0
OpenHermes 2.5	varies per upstream (see dataset card)
Databricks Dolly 15k	CC-BY-SA 3.0

Redistribute accordingly. This dataset is intended for research and pretraining of open language models.

Known limitations

• English-dominant. Some minority coverage in code/arxiv, but this is not a multilingual corpus.
• No PII scrubbing or toxicity filtering beyond what upstream sources applied.
• Upstream deduplication is assumed; no cross-source MinHash dedup is performed here.