bytepft 0.1.0__tar.gz

bytepft-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,27 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  test:
+    # macOS wheels keep the CPU test install free of CUDA runtime dependencies.
+    runs-on: macos-15
+    timeout-minutes: 10
+    env:
+      UV_PYTHON: "3.13"
+      HF_HUB_OFFLINE: "1"
+      OMP_NUM_THREADS: "2"
+    steps:
+      - uses: actions/checkout@v7
+      - uses: astral-sh/setup-uv@v7
+        with:
+          python-version: "3.13"
+      - run: uv sync --frozen
+      - run: uv run --frozen pytest
+      - run: uv run --frozen ruff check .

bytepft-0.1.0/.github/workflows/publish.yml

@@ -0,0 +1,76 @@
+name: Publish to PyPI
+
+on:
+  release:
+    types: [published]
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: Version to build without publishing (for example, v0.1.0)
+        required: true
+        default: v0.1.0
+        type: string
+
+permissions:
+  contents: read
+
+concurrency:
+  group: pypi-${{ github.event.release.tag_name || inputs.tag }}
+  cancel-in-progress: false
+
+jobs:
+  build:
+    runs-on: macos-15
+    timeout-minutes: 15
+    env:
+      RELEASE_TAG: ${{ github.event.release.tag_name || inputs.tag }}
+      UV_PYTHON: "3.13"
+      HF_HUB_OFFLINE: "1"
+      OMP_NUM_THREADS: "2"
+    steps:
+      - uses: actions/checkout@v7
+        with:
+          persist-credentials: false
+      - uses: astral-sh/setup-uv@v7
+        with:
+          python-version: "3.13"
+      - run: uv sync --frozen
+      - name: Set package version from release tag
+        id: prepare
+        run: uv run --frozen python scripts/prepare_release.py
+      - run: uv sync --frozen
+      - name: Verify installed version
+        env:
+          EXPECTED_VERSION: ${{ steps.prepare.outputs.version }}
+        run: |
+          uv run --frozen python - <<'PY'
+          import os
+          from importlib.metadata import version
+          from bytepft import __version__
+          assert version("bytepft") == __version__ == os.environ["EXPECTED_VERSION"]
+          PY
+      - run: uv run --frozen pytest
+      - run: uv build --no-sources
+      - run: uvx twine check dist/*
+      - uses: actions/upload-artifact@v4
+        with:
+          name: release-distributions
+          path: dist/*
+          if-no-files-found: error
+
+  publish:
+    if: github.event_name == 'release'
+    needs: build
+    runs-on: ubuntu-latest
+    environment:
+      name: pypi
+      url: https://pypi.org/project/bytepft/
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: release-distributions
+          path: dist/
+      - name: Publish with PyPI Trusted Publishing
+        uses: pypa/gh-action-pypi-publish@release/v1

bytepft-0.1.0/.gitignore

@@ -0,0 +1,10 @@
+.venv/
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+dist/
+outputs/
+*.egg-info/
+.DS_Store

bytepft-0.1.0/CHANGELOG.md

@@ -0,0 +1,43 @@
+# Changelog
+
+## 0.1.0 — 2026-11-06
+
+Initial experimental release of the BytePFT Python library.
+
+### Added
+
+- Fixed UTF-8 byte vocabulary with separate controls, raw-byte support, batching,
+  tokenizer persistence and incremental text decoding.
+- Causal GRU byte encoder/decoder, fixed/learned/oracle patching, last/mean
+  pooling, causal depooling and a trainable byte/control prediction head.
+- Dense Qwen3 adapter using Hugging Face transformer layers, explicit
+  configuration checks, frozen/global training control and native KV caches.
+- Offline tiny reference workflow and a real 0.6B base-checkpoint CPU smoke test.
+- Diagnostic and source-aligned Stage 1 objectives; explicit optional Stage 2;
+  validation, gradient accumulation/clipping, learning-rate scheduling, early
+  stopping and resumable stage checkpoints.
+- Greedy and sampled byte generation with cached and uncached reference paths.
+- Safetensors model persistence, offline loading and training-state checkpoints.
+- Byte-normalized evaluation, measured generation benchmarking, CLI commands,
+  examples, CPU tests and release tooling.
+
+### Validation scope
+
+CPU tests demonstrate tiny-corpus loss reduction, frozen/unfrozen parameter
+behavior, aligned scalar event losses, deterministic resumption and persistence.
+The real pretrained base checkpoint passes one local gradient update and an exact
+save/load logit comparison. That smoke experiment is **not full adaptation** and
+does not establish retained source-model quality.
+
+### Known limitations
+
+- The reference GRU, causal boundary and separate-head architecture differs from
+  the cited Nature method; exact mLSTM/fused-boundary support is not implemented.
+- Full pretrained adaptation, downstream task quality and matched source-model
+  performance comparisons have not been executed.
+- MPS fp32 passed bounded training/inference and real pretrained integration
+  checks on an M3 Pro. CUDA and MPS mixed precision remain unvalidated. CPU is the default.
+- Inference supports a single generation stream; batched generation and
+  distributed training are not implemented.
+- Source embedding/output matrices are retained in checkpoints for Stage 1;
+  inference does not use the source tokenizer.

bytepft-0.1.0/CONTRIBUTING.md

@@ -0,0 +1,84 @@
+# Contributing to BytePFT
+
+BytePFT prioritizes a small, understandable and verifiable byte-model workflow.
+Contributions should preserve causal alignment, document boundaries and the
+distinction between reference baselines and research reproductions.
+
+## Development setup
+
+Use Python 3.11+ and `uv`:
+
+```sh
+uv sync
+uv run pytest
+uv run ruff check .
+uv run mypy
+uv build
+```
+
+The ordinary suite is CPU compatible and must not download pretrained weights.
+Use `uv sync --extra training` when working on the optional Hugging Face dataset
+integration. Tests use tiny randomly initialized Qwen3 configurations to exercise
+the real Hugging Face transformer implementation.
+
+Defaults use CPU. Do not introduce automatic GPU execution into tests or
+examples. CUDA/MPS tests and pretrained-weight downloads must be explicit opt-ins
+and describe their hardware and download requirements. A contributor must choose
+to run those experiments; CI should remain usable without them.
+
+## Making a change
+
+Keep changes focused, match the surrounding style and add tests for observable
+behavior. Test the failure path when adding validation or serialization code.
+For neural components, cover shapes, masks, finite gradients and relevant causal
+invariants. Changing future bytes must not change predictions for earlier bytes.
+Patch construction must account for every valid byte once, in order.
+
+For training changes, check that Stage 1 leaves global parameters unchanged and
+that optional Stage 2 unfreezing updates them. Resume tests should compare the
+same training configuration and data against a continuous run, including optimizer,
+scheduler, random sampling state and history. For inference changes, compare
+cached logits or greedy bytes against the uncached reference implementation.
+
+Run the tests appropriate to the change, then the ordinary suite and tooling
+before proposing a release. Report the commands actually run and any unavailable
+validation. Avoid claiming that a successful forward pass, random-weight test or
+single gradient step proves model quality.
+
+## Research and API changes
+
+Read the relevant methods and upstream implementation before adding a derived
+architecture or objective. Record the paper section, source revision, deviations
+and validation in [docs/research.md](docs/research.md). Preserve applicable licenses
+when incorporating third-party code. Original additions use Apache-2.0.
+
+Compare teacher and student distributions only when their events and axes are
+mathematically aligned. Source-token logits and byte logits cannot be directly
+KL-diverged. Document byte offsets, BOS/EOS treatment and the prediction shift
+when adding an alignment method.
+
+Do not expose an unfinished adapter or command as supported. Add model-family
+compatibility checks rather than relying on coincidentally similar module paths.
+The current exact-paper mLSTM and fused-boundary architecture remain unsupported.
+
+## Reporting an issue
+
+Include a minimal example, Python/PyTorch/Transformers versions, device and dtype,
+configuration, traceback and whether the model is random, source pretrained or a
+trained BytePFT checkpoint. For resume issues, include the checkpoint stage/step
+and whether the config or documents changed. Never include credentials or private
+training data in a public report.
+
+## Release checklist
+
+1. Update [CHANGELOG.md](CHANGELOG.md) and choose a release tag such as `v0.1.0`.
+2. Run the ordinary CPU suite, Ruff, mypy and `uv build`.
+3. Inspect wheel/sdist contents and install the wheel into a clean environment.
+4. Review [STATUS.md](STATUS.md), examples and measured results for accurate claims.
+5. Complete the one-time [PyPI setup](docs/releases.md), then publish a GitHub
+   release from the tested commit. The release tag sets the package version and
+   triggers publication automatically. A manual workflow dispatch builds and
+   tests a candidate without publishing. Ordinary pushes and pull requests do
+   not publish packages.
+
+Do not add fabricated authorship or automatic co-author trailers to commits.

bytepft-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        https://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "{}"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright {yyyy} {name of copyright owner}
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       https://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

bytepft-0.1.0/NOTICE

@@ -0,0 +1,8 @@
+BytePFT
+Copyright 2026 BytePFT contributors
+
+Original BytePFT code is licensed under Apache-2.0.
+Research attribution: Minixhofer et al., Retrofitting language models to operate over bytes,
+Nature (2026), https://www.nature.com/articles/s41586-026-11111-4
+BytePFT is an independent reference implementation of the cited research.
+Qwen model weights and Hugging Face dependencies retain their respective licenses.

bytepft-0.1.0/PKG-INFO

@@ -0,0 +1,119 @@
+Metadata-Version: 2.5
+Name: bytepft
+Version: 0.1.0
+Summary: Retrofitting language models, one byte at a time.
+Project-URL: Repository, https://github.com/skundu42/bytepft
+Project-URL: Issues, https://github.com/skundu42/bytepft/issues
+Project-URL: Documentation, https://github.com/skundu42/bytepft/tree/main/docs
+Author: BytePFT contributors
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Keywords: distillation,language-model,pytorch,utf-8
+Classifier: Development Status :: 3 - Alpha
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.11
+Requires-Dist: pyyaml>=6
+Requires-Dist: safetensors>=0.5
+Requires-Dist: torch<3,>=2.6
+Requires-Dist: transformers<5,>=4.57.6
+Provides-Extra: benchmark
+Requires-Dist: psutil>=6; extra == 'benchmark'
+Provides-Extra: training
+Requires-Dist: accelerate>=1; extra == 'training'
+Requires-Dist: datasets>=3; extra == 'training'
+Description-Content-Type: text/markdown
+
+# BytePFT
+
+[![CI](https://github.com/skundu42/bytepft/actions/workflows/ci.yml/badge.svg)](https://github.com/skundu42/bytepft/actions/workflows/ci.yml)
+[![Python](https://img.shields.io/badge/python-3.11%2B-blue)](https://www.python.org/downloads/)
+[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/skundu42/bytepft/blob/main/LICENSE)
+
+**Retrofitting language models, one byte at a time.**
+
+BytePFT is a Python library for adapting pretrained Qwen3 language models to
+UTF-8 bytes. It trains a byte-level encoder and decoder around the existing
+transformer, with patch-based processing and cached byte generation.
+
+BytePFT is experimental. The reference implementation is tested on CPU and Apple
+MPS in fp32; full pretrained adaptation and retained language quality have not yet
+been evaluated. See [project status](https://github.com/skundu42/bytepft/blob/main/STATUS.md) for results and limitations.
+
+## Installation
+
+Requires Python 3.11+.
+
+```sh
+pip install bytepft
+```
+
+Or install from source:
+
+```sh
+pip install git+https://github.com/skundu42/bytepft.git
+```
+
+## Quickstart
+
+Train a tiny model locally, without downloading pretrained weights:
+
+```python
+from bytepft import BytePFT, BytePFTConfig
+
+model = BytePFT.from_tiny(
+    config=BytePFTConfig(hidden_size=32, encoder_layers=1, decoder_layers=1),
+    device="cpu",
+)
+model.retrofit(
+    ["hello byte world\n"] * 4,
+    stage1_steps=100,
+    learning_rate=0.01,
+    validation_fraction=0,
+    output_dir="outputs/tiny",
+)
+
+model = BytePFT.load("outputs/tiny")
+print(model.generate("hello", max_new_bytes=32, temperature=0))
+```
+
+This example uses a randomly initialized tiny transformer. To adapt pretrained
+weights, use `BytePFT.from_pretrained("Qwen/Qwen3-0.6B-Base")` and follow the
+[training guide](https://github.com/skundu42/bytepft/blob/main/docs/training.md). Adaptation must complete before generation.
+
+## Features
+
+- UTF-8 byte encoding with separate padding, beginning and end symbols.
+- Fixed, oracle and learned patch boundaries.
+- Frozen-backbone Stage 1 and optional end-to-end Stage 2 training.
+- Resumable checkpoints, validation and bits-per-byte evaluation.
+- Greedy and sampled generation with local state and global KV caching.
+
+Defaults use CPU. Select `device="mps"` explicitly for Apple GPU execution;
+[MPS validation](https://github.com/skundu42/bytepft/blob/main/docs/mps.md) currently covers fp32. CUDA remains unvalidated.
+
+## Command line
+
+```sh
+bytepft train-tiny --output outputs/tiny --device cpu
+bytepft generate --model outputs/tiny --prompt "hello" --max-new-bytes 32
+bytepft inspect Qwen/Qwen3-0.6B-Base
+```
+
+Run `bytepft --help` for all commands.
+
+## Documentation
+
+- [Training and checkpoints](https://github.com/skundu42/bytepft/blob/main/docs/training.md)
+- [Architecture](https://github.com/skundu42/bytepft/blob/main/docs/architecture.md)
+- [Research notes](https://github.com/skundu42/bytepft/blob/main/docs/research.md)
+- [Examples](https://github.com/skundu42/bytepft/tree/main/examples)
+- [Contributing](https://github.com/skundu42/bytepft/blob/main/CONTRIBUTING.md) · [Releases](https://github.com/skundu42/bytepft/blob/main/docs/releases.md) · [Changelog](https://github.com/skundu42/bytepft/blob/main/CHANGELOG.md)
+
+Inspired by [*Retrofitting language models to operate over bytes*, Nature (2026)](https://www.nature.com/articles/s41586-026-11111-4).
+BytePFT uses a simplified reference architecture; the research notes document its deviations.
+
+## License
+
+[Apache-2.0](https://github.com/skundu42/bytepft/blob/main/LICENSE). See [NOTICE](https://github.com/skundu42/bytepft/blob/main/NOTICE) for attribution.

bytepft-0.1.0/README.md

@@ -0,0 +1,92 @@
+# BytePFT
+
+[![CI](https://github.com/skundu42/bytepft/actions/workflows/ci.yml/badge.svg)](https://github.com/skundu42/bytepft/actions/workflows/ci.yml)
+[![Python](https://img.shields.io/badge/python-3.11%2B-blue)](https://www.python.org/downloads/)
+[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/skundu42/bytepft/blob/main/LICENSE)
+
+**Retrofitting language models, one byte at a time.**
+
+BytePFT is a Python library for adapting pretrained Qwen3 language models to
+UTF-8 bytes. It trains a byte-level encoder and decoder around the existing
+transformer, with patch-based processing and cached byte generation.
+
+BytePFT is experimental. The reference implementation is tested on CPU and Apple
+MPS in fp32; full pretrained adaptation and retained language quality have not yet
+been evaluated. See [project status](https://github.com/skundu42/bytepft/blob/main/STATUS.md) for results and limitations.
+
+## Installation
+
+Requires Python 3.11+.
+
+```sh
+pip install bytepft
+```
+
+Or install from source:
+
+```sh
+pip install git+https://github.com/skundu42/bytepft.git
+```
+
+## Quickstart
+
+Train a tiny model locally, without downloading pretrained weights:
+
+```python
+from bytepft import BytePFT, BytePFTConfig
+
+model = BytePFT.from_tiny(
+    config=BytePFTConfig(hidden_size=32, encoder_layers=1, decoder_layers=1),
+    device="cpu",
+)
+model.retrofit(
+    ["hello byte world\n"] * 4,
+    stage1_steps=100,
+    learning_rate=0.01,
+    validation_fraction=0,
+    output_dir="outputs/tiny",
+)
+
+model = BytePFT.load("outputs/tiny")
+print(model.generate("hello", max_new_bytes=32, temperature=0))
+```
+
+This example uses a randomly initialized tiny transformer. To adapt pretrained
+weights, use `BytePFT.from_pretrained("Qwen/Qwen3-0.6B-Base")` and follow the
+[training guide](https://github.com/skundu42/bytepft/blob/main/docs/training.md). Adaptation must complete before generation.
+
+## Features
+
+- UTF-8 byte encoding with separate padding, beginning and end symbols.
+- Fixed, oracle and learned patch boundaries.
+- Frozen-backbone Stage 1 and optional end-to-end Stage 2 training.
+- Resumable checkpoints, validation and bits-per-byte evaluation.
+- Greedy and sampled generation with local state and global KV caching.
+
+Defaults use CPU. Select `device="mps"` explicitly for Apple GPU execution;
+[MPS validation](https://github.com/skundu42/bytepft/blob/main/docs/mps.md) currently covers fp32. CUDA remains unvalidated.
+
+## Command line
+
+```sh
+bytepft train-tiny --output outputs/tiny --device cpu
+bytepft generate --model outputs/tiny --prompt "hello" --max-new-bytes 32
+bytepft inspect Qwen/Qwen3-0.6B-Base
+```
+
+Run `bytepft --help` for all commands.
+
+## Documentation
+
+- [Training and checkpoints](https://github.com/skundu42/bytepft/blob/main/docs/training.md)
+- [Architecture](https://github.com/skundu42/bytepft/blob/main/docs/architecture.md)
+- [Research notes](https://github.com/skundu42/bytepft/blob/main/docs/research.md)
+- [Examples](https://github.com/skundu42/bytepft/tree/main/examples)
+- [Contributing](https://github.com/skundu42/bytepft/blob/main/CONTRIBUTING.md) · [Releases](https://github.com/skundu42/bytepft/blob/main/docs/releases.md) · [Changelog](https://github.com/skundu42/bytepft/blob/main/CHANGELOG.md)
+
+Inspired by [*Retrofitting language models to operate over bytes*, Nature (2026)](https://www.nature.com/articles/s41586-026-11111-4).
+BytePFT uses a simplified reference architecture; the research notes document its deviations.
+
+## License
+
+[Apache-2.0](https://github.com/skundu42/bytepft/blob/main/LICENSE). See [NOTICE](https://github.com/skundu42/bytepft/blob/main/NOTICE) for attribution.