viniciusxpb/scaffold-tokens-dataset

Scaffold Tokens Dataset

Pre-tokenized Portuguese news articles with scaffold (countdown) tokens for training language models with perfect length control.

Resources

Resource	Link
Pre-trained model	viniciusxpb/scaffold-gpt2-pt
Training code	github.com/viniciusxpb/scaffold-tokens
Author	Vinícius França

What are Scaffold Tokens?

Each word in the text is preceded by a countdown token <ff_N> that tells the model how many words remain until the end of the document:

<ff_6> O <ff_5> presidente <ff_4> anunciou <ff_3> novas <ff_2> medidas <ff_1> econômicas <ff_0> .

The model learns to use this signal for exact length control and emergent structural planning — starting with introductory language when ff is high and shifting to conclusions when ff approaches zero.

Original Text Format (JSON)

Before being converted to binary shards, each article is stored as JSON with the countdown annotations:

{
  "id": 1,
  "content-ff": "<ff_630> Com <ff_629> a <ff_628> possibilidade <ff_627> de <ff_626> uma <ff_625> condenação ... <ff_1> financeiro <ff_0> ."
}

During shard creation, each <ff_N> is converted to a numeric token ID (50257 + N) and each word is encoded with GPT-2 BPE.

Dataset Details

Field	Value
Source	Folha de S.Paulo news articles (public domain)
Language	pt-BR
Total tokens	~208M
Tokenizer	tiktoken GPT-2 (50,257 BPE tokens)
Format	Binary shards (uint16)
License	CC0 (Public Domain)

Files

data/
├── train/
│   ├── shard_000.bin   (191 MB, 100M tokens)
│   └── shard_001.bin   (186 MB, 97M tokens)
└── val/
    └── shard_000.bin   (20 MB, 10M tokens)

Shard Format

Each .bin file is a flat array of uint16 values with a header:

• Header: 256 × int32 values
  • [0] = magic number (20240520)
  • [1] = version (1)
  • [2] = token count
• Body: uint16 token IDs

Token Layout

IDs 0–50256:       Standard GPT-2 BPE tokens (tiktoken)
ID  50256:         EOT (end of text, separates documents)
IDs 50257–51256:   <ff_0> through <ff_999> (countdown tokens)

Pattern within documents

[ff_id] [bpe_tok...] [ff_id] [bpe_tok...] ... [EOT]

Each ff token precedes one word. Since Portuguese words often split into multiple BPE subwords, the spacing between ff tokens is variable (1–4+ BPE tokens per word).

How to Load

import numpy as np

def load_shard(path):
    header = np.fromfile(path, dtype=np.int32, count=256)
    assert header[0] == 20240520, "Invalid magic number"
    token_count = header[2]
    tokens = np.fromfile(path, dtype=np.uint16, offset=1024)
    assert len(tokens) == token_count
    return tokens

How to Decode

import tiktoken

enc = tiktoken.get_encoding("gpt2")
FF_BASE = 50257

def decode_shard(tokens):
    words = []
    current_word_tokens = []
    for t in tokens:
        if FF_BASE <= t <= 51256:
            if current_word_tokens:
                words.append(enc.decode(current_word_tokens))
                current_word_tokens = []
        elif t == 50256:  # EOT
            if current_word_tokens:
                words.append(enc.decode(current_word_tokens))
                current_word_tokens = []
            words.append("\n\n")
        else:
            current_word_tokens.append(t)
    return " ".join(words)

Quick Start

# Clone the training repo
git clone https://github.com/viniciusxpb/scaffold-tokens
cd scaffold-tokens

# Setup and download
make setup
make download

# Validate and train
make validate
make train

Citation

@misc{scaffold-tokens-2025,
  title={Scaffold Tokens: Teaching LLMs to Plan with Countdown Tokens},
  author={Vinícius França},
  year={2025},
  url={https://github.com/viniciusxpb/scaffold-tokens}
}