tokenizerbench 0.2.0__tar.gz

tokenizerbench-0.2.0/.github/workflows/python-publish.yml

@@ -0,0 +1,70 @@
+# This workflow will upload a Python Package to PyPI when a release is created
+# For more information see: https://docs.github.com/en/actions/automating-builds-and-tests/building-and-testing-python#publishing-to-package-registries
+
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+name: Upload Python Package
+
+on:
+  release:
+    types: [published]
+
+permissions:
+  contents: read
+
+jobs:
+  release-build:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.x"
+
+      - name: Build release distributions
+        run: |
+          # NOTE: put your own distribution build steps here.
+          python -m pip install build
+          python -m build
+
+      - name: Upload distributions
+        uses: actions/upload-artifact@v4
+        with:
+          name: release-dists
+          path: dist/
+
+  pypi-publish:
+    runs-on: ubuntu-latest
+    needs:
+      - release-build
+    permissions:
+      # IMPORTANT: this permission is mandatory for trusted publishing
+      id-token: write
+
+    # Dedicated environments with protections for publishing are strongly recommended.
+    # For more information, see: https://docs.github.com/en/actions/deployment/targeting-different-environments/using-environments-for-deployment#deployment-protection-rules
+    environment:
+      name: pypi
+      # OPTIONAL: uncomment and update to include your PyPI project URL in the deployment status:
+      # url: https://pypi.org/p/YOURPROJECT
+      #
+      # ALTERNATIVE: if your GitHub Release name is the PyPI project version string
+      # ALTERNATIVE: exactly, uncomment the following line instead:
+      # url: https://pypi.org/project/YOURPROJECT/${{ github.event.release.name }}
+
+    steps:
+      - name: Retrieve release distributions
+        uses: actions/download-artifact@v4
+        with:
+          name: release-dists
+          path: dist/
+
+      - name: Publish release distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          packages-dir: dist/

tokenizerbench-0.2.0/.gitignore

@@ -0,0 +1,212 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[codz]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py.cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# UV
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#uv.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+#poetry.toml
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#   pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
+#   https://pdm-project.org/en/latest/usage/project/#working-with-version-control
+#pdm.lock
+#pdm.toml
+.pdm-python
+.pdm-build/
+
+# pixi
+#   Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
+#pixi.lock
+#   Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
+#   in the .venv directory. It is recommended not to include this directory in version control.
+.pixi
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.envrc
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+# Abstra
+# Abstra is an AI-powered process automation framework.
+# Ignore directories containing user credentials, local state, and settings.
+# Learn more at https://abstra.io/docs
+.abstra/
+
+# Visual Studio Code
+#  Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore 
+#  that can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
+#  and can be added to the global gitignore or merged into this file. However, if you prefer, 
+#  you could uncomment the following to ignore the entire vscode folder
+# .vscode/
+
+# Ruff stuff:
+.ruff_cache/
+
+# PyPI configuration file
+.pypirc
+
+# Cursor
+#  Cursor is an AI-powered code editor. `.cursorignore` specifies files/directories to
+#  exclude from AI features like autocomplete and code analysis. Recommended for sensitive data
+#  refer to https://docs.cursor.com/context/ignore-files
+.cursorignore
+.cursorindexingignore
+
+# Marimo
+marimo/_static/
+marimo/_lsp/
+__marimo__/
+
+
+# Code files
+rough.py
+test.py

tokenizerbench-0.2.0/CHANGELOG.md

@@ -0,0 +1,254 @@
+# TokenizerBench — Release Notes
+
+## v0.2.0
+
+**The "Actually Useful" Release** — this version transforms TokenizerBench from a thin proof-of-concept into a production-ready benchmark you can run before committing to a tokenizer for pre-training or fine-tuning.
+
+---
+
+### What's new at a glance
+
+| | v0.1.0 | v0.2.0 |
+|---|---|---|
+| Human languages | ~13 | **84** |
+| Samples per language | 4 | **20** |
+| Total samples | ~200 | **1,967** |
+| Metrics | none | **8 metrics** |
+| CLI runner | none | ✅ |
+| Tests | none | **101 passing** |
+| CI | none | ✅ GitHub Actions |
+| Analysis / plotting | none | ✅ |
+
+---
+
+### Breaking changes
+
+**Import paths have changed.** All data modules now export by their semantic name instead of a generic `dataset` variable:
+
+```python
+# v0.1.0 — broken
+from data.human_languages import dataset
+
+# v0.2.0 — correct
+from data import human_languages, programming_languages, scientific_formulas, edge_cases
+from data import ALL_DATA  # combined dict of all four
+```
+
+This matches the examples in the README and is required for the new evaluation pipeline to work.
+
+---
+
+### Dataset
+
+#### Human languages — 84 languages, 1,193 samples
+
+The human language dataset has grown from a handful of major languages to **84 languages** covering every major script and language family. Each language now has **20 samples** instead of 4, specifically designed to stress-test tokenizers:
+
+- Standard sentences
+- Long sequences (50–100 words)
+- Repetition stress tests (`word ` × 40)
+- Mixed-script / code-switching samples
+- Punctuation and numeral variants
+- Uppercase-only sentences
+- Script alphabet listings
+
+**New languages added in this release:**
+
+*Indic scripts:* Punjabi, Gujarati, Marathi, Telugu, Kannada, Malayalam, Odia
+
+*East Asian:* Chinese (Traditional)
+
+*Slavic:* Bulgarian, Belarusian, Azerbaijani, Kazakh, Uzbek
+
+*African:* Hausa, Yoruba, Igbo, Somali, Xhosa, Shona
+
+*Southeast Asian:* Tagalog, Cebuano, Javanese
+
+*European minority / constructed:* Catalan, Latvian, Estonian, Slovenian, Luxembourgish, Maltese, Occitan, Scots Gaelic, Esperanto, Latin
+
+*Middle Eastern:* Pashto, Sindhi, Uyghur
+
+#### Programming languages — 17 subcategories, 245 samples
+
+All existing subcategories expanded from 4 to 15–20 samples. Two new subcategories:
+
+- **`typescript`** — generics, decorators, mapped types, `satisfies`, conditional types, `declare module`
+- **`python_ml`** — HuggingFace Transformers, PyTorch, `accelerate`, `datasets`, mixed-precision training code
+
+#### Scientific formulas — 13 subcategories, 237 samples
+
+Three new subcategories:
+
+- **`maxwell_equations`** — all four Maxwell equations in differential form, wave equations, Poynting vector, field energy density
+- **`machine_learning`** — cross-entropy loss, gradient descent, backprop, attention, softmax, BatchNorm, Adam, LoRA, REINFORCE, Bellman, actor-critic, Transformer FFN
+- **`latex_notation`** — raw LaTeX strings with backslash commands, fractions, integrals, matrix notation, `\mathcal`, `\mathbf`, `\nabla`, `\text{}`
+
+#### Edge cases — NEW, 16 subcategories, 292 samples
+
+An entirely new dataset category targeting known tokenizer failure modes:
+
+| Subcategory | What it catches |
+|---|---|
+| `homoglyphs` | Cyrillic/Greek/Latin lookalikes silently mixed into text |
+| `zero_width_characters` | ZWS, ZWNJ, ZWJ, BOM, soft hyphens changing token boundaries |
+| `rtl_ltr_mixing` | Arabic/Hebrew + Latin bidirectional text |
+| `diacritics_and_special_latin` | Combining vs precomposed forms (NFC/NFD equivalence) |
+| `whitespace_variants` | All 15 Unicode space types, tabs, various newlines |
+| `long_tokens` | URLs, base64, UUIDs, SHA hashes, long identifiers |
+| `repeated_characters` | Single-char spam, emoji runs, repeated n-grams |
+| `emojis_and_unicode` | Flag sequences, ZWJ family emoji, skin tone modifiers, keycaps |
+| `code_switching` | Natural language sentences with embedded code keywords |
+| `noisy_text` | OCR noise, missing spaces, typos, leetspeak, alternating case |
+| `mixed_scripts_single_token` | Diacritic words that should tokenize as a single unit |
+| `numerical_edge_cases` | Hex, binary, fractions, Roman numerals, Unicode digits |
+| `special_punctuation` | 6 quote styles, 5 dash types, arrows, all currency symbols |
+| `control_characters` | Null bytes, ANSI escape codes, Unicode control chars |
+| `fertility_test` | Morphologically complex words (Turkish, Finnish, Polish, Tamil…) |
+| `segmentation_boundaries` | Contractions, possessives, hyphenation, abbreviations |
+
+---
+
+### Metrics (`metrics.py`)
+
+A new module implementing 8 evaluation metrics that work with any tokenizer implementing `.encode(text) → list[int]` and `.decode(ids) → str`:
+
+| Function | Description |
+|---|---|
+| `token_count` | Raw token count |
+| `fertility_score` | Tokens per word — the standard multilingual quality metric |
+| `compression_ratio` | Tokens per character |
+| `byte_compression_ratio` | Tokens per UTF-8 byte — language-agnostic |
+| `roundtrip_fidelity` | Whether `decode(encode(text)) == text` |
+| `vocabulary_coverage` | Fraction of IDs within the known vocabulary range |
+| `subword_consistency` | Consistency of morphological root segmentation |
+| `segmentation_stats` | Min/max/mean/std/p90/p95/p99 of token counts |
+
+High-level pipelines:
+
+- `evaluate_tokenizer(tokenizer, dataset)` — runs all metrics across every category and subcategory, returns nested results dict with a `__summary__` block
+- `compare_tokenizers(tokenizers_dict, dataset)` — evaluates multiple tokenizers in one call
+- `make_leaderboard(comparison)` — returns a sorted ranking list
+
+#### Fertility score interpretation
+
+```
+≈ 1.0   ideal — one token per word
+1 – 2   good  — well-trained BPE on a covered language
+2 – 4   acceptable — less common scripts
+≥ 4     poor  — language likely under-represented in training data
+```
+
+#### Roundtrip fidelity
+
+Any fidelity failure is a **bug** in the tokenizer. The metric returns the exact character position of the first divergence, making it easy to bisect which input triggered the failure. Run the `edge_cases` dataset to catch these before training.
+
+---
+
+### CLI runner (`evaluate.py`)
+
+```bash
+# Single tokenizer
+python evaluate.py --tokenizer tiktoken --model cl100k_base
+
+# HuggingFace
+python evaluate.py --tokenizer hf --model xlm-roberta-base
+
+# SentencePiece
+python evaluate.py --tokenizer sentencepiece --model spm.model
+
+# Subset of categories
+python evaluate.py --tokenizer tiktoken --categories human_languages edge_cases
+
+# Save to JSON
+python evaluate.py --tokenizer tiktoken --output results/cl100k.json
+
+# Side-by-side comparison + leaderboard
+python evaluate.py \
+  --tokenizer tiktoken tiktoken hf \
+  --model cl100k_base o200k_base xlm-roberta-base
+```
+
+Supported backends: `tiktoken`, `hf` / `huggingface`, `sentencepiece` / `sp`.
+
+---
+
+### Tests (`tests/`)
+
+101 tests, zero external dependencies required.
+
+- **`tests/conftest.py`** — `MockTokenizer` (char-level, pure stdlib), `BrokenDecodeTokenizer` (always fails fidelity), `SpaceTokenizer` (word-level baseline)
+- **`tests/test_metrics.py`** — 56 unit tests, one per metric function and edge case
+- **`tests/test_data.py`** — 45 dataset integrity tests: structure, encoding validity, no surrogates, minimum sample counts, content spot-checks
+
+```
+101 passed in 0.21s
+```
+
+---
+
+### CI (`.github/workflows/ci.yml`)
+
+Three jobs on every push and pull request to `main`:
+
+1. **`test`** — pytest on Python 3.10, 3.11, 3.12 (matrix)
+2. **`test-tiktoken`** — integration test with real tiktoken installed
+3. **`lint`** — ruff with E, W, F rules
+
+---
+
+### Analysis (`analysis/plot.py`)
+
+```bash
+# Single tokenizer — generates heatmap, bar chart, scatter plot
+python analysis/plot.py results/cl100k.json --out figures/
+
+# Multi-tokenizer comparison
+python analysis/plot.py results/cl100k.json results/xlm.json --compare
+```
+
+Plots generated:
+
+- **Fertility heatmap** — category × subcategory grid, colour-coded by quality
+- **Per-language bar chart** — all 84 languages sorted by fertility, green/orange/red thresholds
+- **Fertility vs compression scatter** — spot outliers by category
+- **Grouped bar comparison** — side-by-side metric comparison across tokenizers
+- **Fidelity failure bar** — which tokenizers have roundtrip bugs
+
+Requires: `pip install matplotlib seaborn pandas`
+
+---
+
+### Bug fixes
+
+- Fixed `SyntaxError` in `edge_cases.py` caused by unescaped low-high quotation marks (`„"`) inside a double-quoted string literal
+- Fixed `SyntaxWarning` for invalid escape sequence `\\/` in leetspeak sample — replaced with a raw string
+
+---
+
+### Upgrade guide
+
+```bash
+pip install -r requirements.txt
+```
+
+Update any imports from:
+```python
+from data.human_languages import dataset
+```
+to:
+```python
+from data import human_languages
+# or
+from data import ALL_DATA
+```
+
+---
+
+### What's next (v0.3.0 ideas)
+
+- Expand to 100 languages (16 remaining)
+- Longer sequences: 2K–10K character samples for context-length stress tests
+- Baseline JSON files for `cl100k_base`, `o200k_base`, `xlm-roberta-base`, `bert-base-multilingual-cased`
+- Moses and NLTK word-tokenize as word-level fertility baselines
+- Plotting notebook (`analysis/explore.ipynb`)
+- PyPI package (`pip install tokenizerbench`)

tokenizerbench-0.2.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 KiteFishAI
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.