Blame for README.md

---
license: other
task_categories:
- text-generation
language:
- en
size_categories:
- 1B<n<10B
pretty_name: Stew-10B
tags:
- pretraining
- gpt2
- tokenized
- pre-tokenized
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.parquet
---

![a4eabb77-0b9b-447b-a344-aaeaeb181821](https://cdn-uploads.huggingface.co/production/uploads/6a5f8c87a38cac087c2c6c05/O1nffJZXpsyWSmhjVhKBs.jpeg)

# Stew-10B

A 10-billion-token pretraining dataset for language models, already tokenized and ready to feed straight into a training loop.

## "Why am I just seeing numbers?"

If you opened the dataset viewer and saw rows that look like this:

```
[464, 2159, 3666, 318, 257, 1332, 286, 262, ...]
```

...that's not a bug. **This dataset doesn't contain text — it contains the token IDs that text was converted into.**

Here's what's going on. Before a language model can learn from text, every word (or piece of a word) has to be mapped to a number using a "tokenizer." The sentence `"The world"` becomes something like `[464, 995]` using the GPT-2 tokenizer. Models don't read letters — they read these integer IDs.

Most datasets on the Hub ship the raw text and leave tokenization to you. That works, but tokenizing 10 billion tokens worth of text takes **hours of CPU time** every time someone starts a training run. It's wasted work, repeated by everyone.

**Stew-10B skips that step.** The text has already been tokenized once, using the standard GPT-2 tokenizer, and the resulting integer IDs are what's stored here. You get:

- **Instant training startup** — no tokenization phase, just load and train
- **Deterministic data** — every user sees the exact same token stream, so experiments are reproducible
- **Smaller download** — token IDs compress better than raw UTF-8 text

The tradeoff is that the data is no longer human-readable in the viewer. If you want to see what a row actually says, you decode it:

```python
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("gpt2")
print(tok.decode([464, 2159, 3666, 318, 257, 1332, 286, 262]))
# -> "The World My is a test of the"
```

## What's actually in each row

Each row in the parquet files contains **one training window of exactly 2,048 tokens** under the column `input_ids`. That's roughly 1,500 words of English text per row, which is a standard context length for pretraining smaller GPT-style models.

Rows are packed back-to-back from the original token stream, with documents separated by the special `<|endoftext|>` marker (token ID `50256`). So within a single row you might find the end of one Wikipedia article, the EOS token, and then the start of a web page — this is the standard "packed sequences" format used by nanoGPT, GPT-NeoX, Megatron, and basically every modern pretraining codebase.

## Where the text came from

Stew-10B is a careful mix of three well-known English pretraining corpora, blended at the token level to hit exact proportions:

| Source | Share | Tokens | What it is |
|---|---|---|---|
| [`wikimedia/wikipedia`](https://huggingface.co/datasets/wikimedia/wikipedia) (`20231101.en`) | 25% | 2.5B | Clean, encyclopedic English — the "high-quality" anchor |
| [`openbmb/Ultra-FineWeb`](https://huggingface.co/datasets/openbmb/Ultra-FineWeb) (English split) | 50% | 5.0B | Filtered Common Crawl — broad coverage of the open web |
| [`mlfoundations/dclm-baseline-1.0`](https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0) | 25% | 2.5B | DataComp-LM baseline — aggressively filtered web text |

The mix is designed to balance **quality** (Wikipedia), **diversity** (FineWeb), and **scale-trained signal** (DCLM) in roughly the proportions that have been shown to work well in recent small-model pretraining recipes.

## How to use it

### Option 1: Hugging Face `datasets` (easy mode)

```python
from datasets import load_dataset

ds = load_dataset("Monster-Code/Stew-10B-tokenized", split="train", streaming=True)

for row in ds:
    tokens = row["input_ids"]   # a list of 2048 integers
    # ...feed into your model
    break
```

Streaming is recommended because the full dataset is around 25 GB and you probably don't want all of it on disk at once.

### Option 2: Raw binary files (fast mode)

If you're running something like nanoGPT and want zero-overhead access, there's a second copy of the data stored as raw `uint16` binary files under `data_bin/`. These are flat token streams you can memory-map directly:

```python
import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "Monster-Code/Stew-10B-tokenized",
    "data_bin/train-00000-of-00020.bin",
    repo_type="dataset",
)
tokens = np.memmap(path, dtype=np.uint16, mode="r")
print(tokens.shape)      # (500_000_000,)
print(tokens[:16])       # first 16 token IDs
```

Each `.bin` file holds 500 million tokens (1 GB), and there are 20 of them. Concatenated in order, they reconstruct the full 10-billion-token stream.

## File layout

```
Monster-Code/Stew-10B-tokenized/
├── data/                               ← the parquet files (what `datasets` loads)
│   ├── train-00000-of-00020.parquet
│   ├── train-00001-of-00020.parquet
│   └── ...  (20 files total)
└── data_bin/                           ← same tokens as raw uint16 binary
    ├── train-00000-of-00020.bin
    ├── train-00001-of-00020.bin
    └── ...  (20 files total)
```

Both folders contain the same tokens in the same order. Pick whichever format fits your pipeline.

## Technical details

- **Tokenizer:** `gpt2` (the fast / Rust implementation from `tokenizers`)
- **Vocabulary size:** 50,257
- **End-of-document marker:** `<|endoftext|>` (token ID `50256`), inserted after every document
- **Sequence length (parquet):** 2,048 tokens per row
- **Total tokens:** 10,000,000,000 (exact — enforced during the build)
- **Total parquet rows:** approximately 4,882,812
- **Parquet schema:** `input_ids: fixed_size_list<int32>[2048]`
- **Parquet compression:** zstd level 3
- **Binary format:** little-endian `uint16`, no header, no padding

## Licensing

This dataset is a derivative work combining three upstream sources. **Each source retains its original license** — you are responsible for complying with all of them if you use this dataset:

- **Wikipedia** → CC BY-SA 3.0 and GFDL ([details](https://huggingface.co/datasets/wikimedia/wikipedia))
- **Ultra-FineWeb** → see the [original dataset card](https://huggingface.co/datasets/openbmb/Ultra-FineWeb)
- **DCLM-baseline** → see the [original dataset card](https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0)

No additional restrictions are imposed by this repository, but the dataset is marked `license: other` because the upstream licenses are mixed and you need to read them yourself before any downstream use.

## Citation

If you use Stew-10B in research, please cite the three underlying datasets. This repository is a repackaging and does not claim original authorship of the text content.