FFBPE

FFBPE is fast and faithful byte-pair encoding for large, multilingual corpora.

Exact BPE at corpus scale.

It combines Unicode-aware inventory shaping, frequency-safe merge cutoffs, exact bounded-memory training, and tiktoken-compatible encoding.

Python is the easiest way to train and use a tokenizer. Rust exposes the lower-level training, encoding, and streaming primitives.

Why FFBPE

• Shape Unicode-heavy inventories before training. A two-pass Unicode-bigram pipeline retains frequent adjacent pairs and splits unproductive boundaries, reducing the nearly unique word inventories common in CJK corpora.
• Carry measured boundaries into training. Bigram selection reports its inclusive frequency cutoff, and BPE training can stop before learning a pair below that boundary. Selection includes all ties at the cutoff.
• Keep rare Unicode scalars encodable without bloating the alphabet. Unicode training can reserve part of its learned vocabulary for byte-level fallback merges inside scalars that were not materialized directly.
• Accelerate long-word encoding when the vocabulary supports it. Encoders can partition PAT words using bigrams already present in the model vocabulary, with an explicit opt-out for workloads where the extra scan is not profitable.
• Bound memory without approximating the model. The optional hot-pair window bounds persistent occurrence postings while preserving global frequencies, winner selection, and deterministic tie-breaking.
• Stream corpora through bounded native batches. Replayable sources, prefetch, mergeable counters, and native counter-to-trainer transfer avoid materializing a complete Python dictionary.
• Use familiar model formats and APIs. Byte models support GPT-2 serialization, Unicode models use the lossless legacy unitoken format, and Python includes a tiktoken-shaped API.

Measured impact

One release run on the 64 MiB FineWeb2 Chinese fixture, training a 10,000-token vocabulary, measured:

For this workload, Unicode-bigram shaping produced an approximately 3× smaller inventory and 7× faster BPE training. It changes corpus segmentation and should be benchmarked on representative text rather than treated as a universal speedup.

On a 1 GiB FineWeb2 Chinese Unicode-bigram inventory, the exact bounded-memory mode reduced observed training peak RSS from 1,797 MiB to 1,649 MiB while producing the same model; training changed from 5.58 s to 5.85 s and required two hydration scans.

See BENCHMARKS.md for benchmark contracts, qualifications, and reproduction commands.

Install

Python 3.11 or newer:

pip install ffbpe

JavaScript (Node or browser, backed by WebAssembly):

npm install @tokn-ai/ffbpe

import { FFBPE, trainBpe } from "@tokn-ai/ffbpe"

await FFBPE.init()
const model = trainBpe("hello tokenizer", { vocab_size: 270 })

Load a common tokenizer lazily with the companion preset package:

npm install @tokn-ai/ffbpe @tokn-ai/ffbpe-presets

import { FFBPE } from "@tokn-ai/ffbpe"
import { loadPreset } from "@tokn-ai/ffbpe-presets"

await FFBPE.init()
const encoder = await loadPreset("cl100k_base")

Presets include only metadata in the npm package. The selected official model asset is downloaded on demand and verified before it is converted in memory.

The npm package uses filesystem paths in Node and File/Blob objects in the browser. See packages/ffbpe/README.md for the full runtime-specific API.

Rust:

cargo add ffbpe

Renamed from unitoken

FFBPE 0.1.8 is the first release under the new package name. Install and import ffbpe for new projects. Existing model directories remain compatible:

• FFBPE writes model metadata to ffbpe.json.
• FFBPE also reads legacy unitoken.json metadata.
• The serialized unitoken model format name remains stable, so existing vocab and merge files do not need conversion.

The old packages will receive a separate deprecation-only release after FFBPE is available.

Five-minute Python quickstart

Train directly from strings, encode and decode without an intermediate file round trip, then save a self-describing model directory:

from ffbpe import BpeEncoder, train_bpe

model = train_bpe(
  ["hello world", "hello tokenizer"],
  vocab_size=280,
  special_tokens=["<|endoftext|>"],
)

ids = model.encode("hello world")
assert model.decode(ids) == "hello world"

model.save_pretrained("my-tokenizer")
encoder = BpeEncoder.from_pretrained("my-tokenizer")
assert encoder.decode(encoder.encode("hello world")) == "hello world"

train_bpe accepts a single string or a one-pass iterable of independent text records. Counting is batched in Rust. Training can finish below the requested size when the corpus has no eligible pairs left. A target below the initial vocabulary size is rejected.

Use BpeTrainer and PreTokenizer directly when you need compressed word counts, Unicode-bigram selection, a custom regex, or manual merge steps.

Encoding partitions long PAT words using model-vocabulary bigrams by default. This does not change token ids. If profiling shows that the scan is slower for your byte model, disable it consistently when creating or saving the encoder:

encoder = model.encoder(split_on_vocab_bigrams=False)
model.save_pretrained("my-tokenizer", split_on_vocab_bigrams=False)

Unicode-bigram training with a safe cutoff

Unicode-bigram shaping is an explicit two-pass workflow:

1. Count adjacent Unicode pairs.
2. Retain the most frequent pairs, including every tie at the boundary.
3. Count words using the retained pairs.
4. Carry the measured cutoff into BPE training.

from ffbpe import BpeTrainer, PreTokenizer


class Corpus:
  def scan(self):
    yield "你好世界"
    yield "你好，tokenizer"


corpus = Corpus()
pretokenizer = PreTokenizer([])

bigram_counter = pretokenizer.bigram_counter()
bigram_counter.add_source(corpus.scan())
selection = bigram_counter.select(top_k=100_000, min_freq=2)

word_counter = (
  pretokenizer
  .with_unicode_bigrams(selection.bigrams)
  .word_counter()
)
word_counter.add_source(corpus.scan())

trainer = BpeTrainer(
  [],
  unit="unicode",
  bigram_cutoff_freq=selection.cutoff_freq,
)
trainer.add_word_counter(word_counter)
trainer.train(vocab_size=10_000)
model = trainer.validate_model()

encoder = model.encoder(unicode_bigrams=selection.bigrams)
model.save_pretrained(
  "my-unicode-tokenizer",
  unicode_bigrams=selection.bigrams,
)

train() stops before a new merge below bigram_cutoff_freq. Equality is valid because selection retains all ties at the cutoff. Manual step() calls remain available, while model validation rejects a final merge below the configured cutoff. Pass the selected bigrams to model.encoder() and model.save_pretrained() as shown; the self-describing directory then restores the same pretokenizer configuration.

unicode_bigram_mixed_boundary="keep" is the conservative default: it preserves mixed or unmeasured edges and splits unretained script-to-script edges. Use "split" only when the more aggressive segmentation matches your intended tokenizer.

Unicode BBPE fallback

Unicode models can spend a configurable share of learned vocabulary slots on byte merges inside rare Unicode scalars:

trainer = BpeTrainer([], unit="unicode")
trainer.add_word_counter(word_counter)
trainer.train_with_bbpe_fallback(
  vocab_size=10_000,
  primary_vocab_ratio=0.9,
)
model = trainer.validate_model()

The ratio applies to learned slots after special tokens and the mandatory 256-byte alphabet. Primary Unicode training runs first; the fallback pass then learns only inside omitted scalars and never across scalar boundaries. Unused fallback slots return to primary training. The resulting model needs no special loading option because the behavior is encoded in its merge rules.

Fallback is a target-aware, finalizing operation and must run before ordinary vocabulary growth. Use primary_vocab_ratio=1.0 when no fallback slots should be reserved; that delegates to ordinary training and leaves the trainer extendable.

Streaming and partitioned counting

add_source pulls at most 4,096 records or 64 MiB per batch by default. It overlaps Python iteration with Rust processing using one bounded look-ahead batch. Pass prefetch=0 for synchronous processing or override max_records and max_bytes for the record sizes and worker memory available.

Counters can be merged after independently counting corpus partitions:

left = pretokenizer.word_counter()
left.add_source(left_partition)

right = pretokenizer.word_counter()
right.add_source(right_partition)

left.merge(right)
trainer.add_word_counter(left)

add_word_counter consumes the native inventory without constructing a Python dictionary; the counter is empty and reusable afterward. word_counter.words() remains available for small inventories but copies the complete result into Python.

Exact bounded-memory training

By default, the trainer retains occurrence postings for every discovered pair. Set hot_pair_window_size to keep an exact top-K candidate window:

trainer = BpeTrainer(
  [],
  unit="unicode",
  hot_pair_window_size=4096,
)
trainer.add_word_counter(word_counter)
trainer.train(vocab_size=10_000)

Smaller windows use less persistent posting memory but may require more full inventory scans when a cold pair wins. Larger windows retain more postings and usually reduce hydration. The setting affects resource use, not pair frequencies or the resulting model. Inspect trainer.hot_pair_window_stats for hydration, pruning, resident-pair, and occurrence-capacity diagnostics.

The hot window is not a total process-memory bound: the full word inventory and a compact global pair table remain resident to preserve exact selection. Inspect trainer.memory_usage for capacity-backed word, pair-table, occurrence, heap, and model storage. Its estimated_persistent_bytes intentionally excludes allocator retention, stacks, and temporary parallel work.

Tiktoken-compatible API

Use the familiar Encoding surface with an existing model:

from ffbpe import Encoding

encoding = Encoding.from_files(
  "my-tokenizer",
  vocab_file="my-tokenizer/vocab.json",
  merges_file="my-tokenizer/merges.txt",
  special_tokens={"<|endoftext|>": 0},
)

ids = encoding.encode("hello world")
text = encoding.decode(ids)

The ffbpe.tiktoken namespace exports Encoding, get_encoding, encoding_for_model, encoding_name_for_model, and list_encoding_names with signatures checked against upstream tiktoken. Built-in registry names are currently limited to fixture models; load trained models explicitly.

Rust quickstart

use ffbpe::{
  bpe::{BpeTrainer, Idx},
  traits::Encode,
};

fn main() -> Result<(), Box<dyn std::error::Error>> {
  let mut trainer = BpeTrainer::<u8, Idx>::from_words(
    [("hello", 10), ("world", 7)],
    &[],
  );
  trainer.train_until(260)?;

  let model = trainer.validate_model()?;
  let encoder = model.to_encoder()?;
  let ids = encoder.encode_string("hello")?;

  assert_eq!(encoder.decode(&ids)?, "hello");
  Ok(())
}

The same program is available as examples/quickstart.rs and compiled in CI.

Known GPT-2, r50k, cl100k, and o200k PAT expressions are implemented by the workspace's ffbpe-pat crate. It can also be used directly when an application needs only zero-copy pretoken ranges or borrowed tokens; arbitrary patterns continue through FFBPE's regex fallback.

The non-default simd feature enables 64-byte boundary masks for ASCII-led GPT-2, r50k, cl100k, and o200k inputs. It uses NEON on AArch64 and selects AVX2 at runtime on x86_64. The portable scanner remains the fallback for short inputs, inputs whose first window contains Unicode, x86_64 CPUs without AVX2, and other architectures.

CLI

The Rust CLI is feature-gated:

cargo run --release --features cli -- train \
  --vocab-size 10000 \
  --out out/models \
  corpus.txt

Run cargo run --features cli -- train --help for chunking, boundary, character-unit, special-token, and output-format options.

Development

Build the Python extension in a local environment:

uv sync --dev
maturin develop --release --features py
uv run pytest

Useful entry points:

• python examples/quickstart.py
• cargo run --example quickstart
• cargo test -p ffbpe-pat
• cargo bench -p ffbpe-pat --bench scan
• python benchmarks/compare_tiktoken.py
• python benchmarks/compare_hf_training.py
• cargo bench --bench regression -- suite smoke
• cargo bench --bench regression -- suite 64mib --check

FFBPE is licensed under the MIT License.