litetune 0.1.0__tar.gz

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

@@ -0,0 +1,50 @@
+name: ci
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+jobs:
+  check:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        # The floor and the current release. The floor is declared in
+        # pyproject.toml and was previously untested, which made it a claim
+        # rather than a fact.
+        python-version: ["3.10", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      - run: pip install -e ".[dev]"
+      - run: ruff check src tests
+      - run: ruff format --check src tests
+      - run: mypy
+      # The whole suite runs on fakes: no subprocess, no network, no GPU.
+      - run: pytest -q
+
+  # An editable install is not the artifact anyone receives. `src/` layout plus
+  # a dynamic hatch version plus `license-files` is exactly the surface where
+  # `pip install -e .` passes and a wheel does not.
+  wheel:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+      - run: pip install build twine
+      - run: python -m build
+      - run: twine check dist/*
+      - name: install the wheel into a clean environment
+        run: |
+          python -m venv /tmp/fresh
+          /tmp/fresh/bin/pip install dist/*.whl
+          /tmp/fresh/bin/litetune --help
+          /tmp/fresh/bin/python -c "import litetune, pathlib; \
+            assert litetune.__version__, 'no version'; \
+            assert (pathlib.Path(litetune.__file__).parent / 'py.typed').is_file(), 'no py.typed'"

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

@@ -0,0 +1,52 @@
+# Publishes to PyPI when a GitHub Release is published.
+#
+# No API token anywhere. PyPI trusts this exact workflow in this exact
+# repository through OIDC, so there is no secret to leak, rotate, or paste into
+# a chat window. `id-token: write` is what mints the short-lived credential;
+# without it the publish step has nothing to authenticate with.
+name: publish
+
+on:
+  release:
+    types: [published]
+
+jobs:
+  pypi:
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      # The whole authentication story. Everything else stays read-only.
+      id-token: write
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      # Imported, not scraped. A regex for the first double-quoted string in
+      # the file matched the module docstring -- which contains `"never raises"`
+      # -- and reported that as the package version. The guard failed safe, by
+      # refusing to publish, but for the wrong reason. `_version.py` is a leaf
+      # module with no imports of its own, so importing it costs nothing.
+      #
+      # The tag is the version. A release built from a tree whose
+      # `_version.py` says something else would publish an artifact whose
+      # recorded version does not lead back to the code that made it -- and
+      # `manifest.py` writes that version into every artifact litetune
+      # produces, so the mismatch would propagate to the user's bundles.
+      - name: The tag and the package must agree
+        run: |
+          set -o pipefail
+          tag="${GITHUB_REF_NAME#v}"
+          pkg=$(PYTHONPATH=src python -c "from litetune._version import __version__; print(__version__)")
+          echo "tag=$tag package=$pkg"
+          test "$tag" = "$pkg" || { echo "::error::tag $tag does not match _version.py $pkg"; exit 1; }
+
+      - name: Build
+        run: |
+          python -m pip install --upgrade build
+          python -m build
+
+      - name: Publish
+        uses: pypa/gh-action-pypi-publish@release/v1

litetune-0.1.0/.gitignore

@@ -0,0 +1,14 @@
+__pycache__/
+*.py[cod]
+.venv/
+venv/
+dist/
+build/
+*.egg-info/
+runs/
+.envs/
+
+# Working notes, not shipped
+openspec/
+.claude/
+.venv/

litetune-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://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 2026 Sasha Denisov
+
+   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
+
+       http://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.

litetune-0.1.0/MEASUREMENTS.md

@@ -0,0 +1,141 @@
+# What was measured, and what it established
+
+Numbers for `litetune`. Everything here is `functiongemma-270m-it` LoRA-tuned on
+`google/mobile-actions`, scored on 640 held-out single-call examples; exact match
+means the tool name **and** every argument value.
+
+This file exists so the README can be a usage guide. It is the longer story:
+what reproduced, what did not, and which published claims were withdrawn.
+
+## The headline numbers
+
+| | float | `dynamic_wi8_afp32` | `weight_only_wi8_afp32` |
+|---|---|---|---|
+| Base model | 0.7266 ±0.0345 | — | — |
+| Fine-tuned | **0.9172** ±0.0214 | 0.9016 ±0.0231 | 0.9047 ±0.0228 |
+| Cost of conversion | — | +0.0156 ±0.0162 *(unresolved)* | **+0.0125** ±0.0123 |
+
+Fine-tuning gained **+0.1906 ±0.0357**, measured paired: only the examples where
+the two models disagree carry signal, and 136 of 640 did.
+
+Pairing is not a refinement for its own sake. `weight_only_wi8_afp32` costs
+**+0.0125**, and the estimate is the same either way — what changes is what can
+be said about it. Paired, the interval is **±0.0123**, just clear of zero, so
+the effect is resolved. Unpaired it is **±0.0312**, which straddles zero and
+would be reported as noise. Both sides answered the same 640 prompts; treating
+them as independent samples throws away exactly the information that settles the
+question.
+
+### The base figure is measured under a rendering the base did not learn
+
+**Read the base column with this in mind.** The declaration properties in every
+prompt above are rendered in *declaration order*, which is what the reference
+consumer (`flutter_gemma`) emits. The jinja template inside the `.litertlm` we
+ship renders them with `dictsort` — measured, eight occurrences, including on
+declaration properties. On this dataset the two orders disagree for **100% of
+rows**, so one bundle presents two different prompts depending on which path a
+consumer takes.
+
+Which order the weights prefer was argued rather than measured until it was
+measured. Same greedy decode, same parser, one variable:
+
+| | n | declaration order | `dictsort` | paired difference | discordant pairs |
+|---|---|---|---|---|---|
+| Base, held-out | 640 | 0.7266 | **0.7625** | −0.0359 ±0.0219 | 51 |
+| Base, disjoint sample | 1280 | 0.7602 | **0.7789** | −0.0187 ±0.0099 | 42 |
+| **Fine-tuned**, held-out | 640 | 0.9109 | 0.9141 | −0.0031 ±0.0087 *(unresolved)* | **8** |
+
+**The base cares and the fine-tuned model does not.** For the base both runs
+resolve, both favour `dictsort`, the intervals overlap at [−0.0286, −0.0140],
+and the discordant pairs run three to one. After fine-tuning the same comparison
+collapses: 51 discordant pairs become 8, and the difference is 0.0031 against an
+interval of ±0.0087 — unresolved not for want of data but for want of an effect.
+
+The mechanism is visible in the failures. It is not a parsing artifact — argument
+dicts compare without regard to key order, so a reordered call scores the same.
+What moves is *which argument the model extracts*: given a declaration in the
+order it did not learn, the base returns `email` where the target wanted
+`phone_number`. It had learned "the Nth property is X". Fine-tuning teaches it to
+read the name instead, and the position stops mattering.
+
+Two consequences for the headline table:
+
+- **The base column is understated**, because it was measured in declaration
+  order. Under the order the weights prefer it is 0.7625.
+- **The fine-tuning gain depends on which base you compare against**, and both
+  are now measured rather than inferred: **+0.1843** entirely in declaration
+  order, **+0.1516** entirely in `dictsort`. The published +0.19 sits above both
+  because it pairs a tuned model measured at its best against a base measured at
+  its worst.
+
+`contract.json` records which convention a bundle was built under, so a consumer
+is not guessing. On the evidence above that matters for a base or lightly-tuned
+checkpoint and is close to free for a fully fine-tuned one — but "close to free"
+is a measurement on one model and one dataset, not a property of the method.
+
+### What three runs of the same thing disagree about
+
+The same recipe sweep, run three times on the same data with the same code. B
+and C are the identical shipping configuration; A differs only in the bundle's
+declared model type, which the text-parsing measurement cannot observe:
+
+| | run A | run B | run C |
+|---|---|---|---|
+| Base, float | 0.7266 | 0.7266 | 0.7266 |
+| Fine-tuned, float | 0.9094 | 0.9172 | 0.9062 |
+| `dynamic_wi8_afp32` | 0.8906 — **resolved** | 0.9016 — *unresolved* | 0.8969 — *unresolved* |
+| `weight_only_wi8_afp32` | 0.9141 — *unresolved* | 0.9047 — **resolved** | 0.9000 — *unresolved* |
+| Gap between recipes | 0.0235 | 0.0031 | 0.0031 |
+
+**The recipes swap places, and which cost resolves moves with them.** An earlier
+draft of this file read "the two recipes differ by 0.0234 on the same weights at
+the same bit width" and drew a conclusion from it. Two further runs put that gap
+at 0.0031 and 0.0031. The gap was noise, and a single run had presented it as a
+finding.
+
+Across all three, the conversion cost resolves in **2 of 6** recipe-runs -- and
+B and C, which are the same configuration end to end, still disagree about
+`weight_only`.
+
+So what is actually established is narrower than one run suggests, and worth
+separating:
+
+- **The base figure is 0.7266**, identical to four decimal places in all three,
+  and reproduced across a rebuilt container image, a `transformers` major version
+  change, new batching code and a rewritten parser.
+- **Fine-tuning gains about +0.18**, resolved in every run, on 132-136 discordant
+  pairs. The spread across runs is 0.0109 — an order below the effect.
+- **Conversion costs something small** — 0.0063 to 0.0187 across runs — and
+  whether 640 examples *resolve* it is close to a coin flip. At this effect size
+  the method is at its limit, and one run's verdict should not be quoted as the
+  answer.
+
+That last line is the tool working, not failing. A single accuracy number would
+have shown none of this; three points and a paired interval show exactly where
+the evidence stops. If you need to separate two recipes this close, you need
+more held-out examples than 640 — and `verify` will keep saying "unresolved"
+until you have them, rather than picking a winner.
+
+Nothing in the file size, the exit code, or the logs separates those two
+artifacts: they are 455,759,152 and 455,939,600 bytes, 0.04% apart. Running both
+against held-out data is the only thing that does.
+
+Three points rather than two, because only the differences mean anything. A
+single accuracy figure for a converted model cannot distinguish a good
+conversion of a bad model from a bad conversion of a good one.
+
+**One `verify` run measures two of them, not three.** It compares the converted
+model against one reference, and which reference you name decides which
+difference you get:
+
+| `--reference-role` | reference is | you get | you do not get |
+|---|---|---|---|
+| `float_twin` (default) | the checkpoint the artifact was converted from | conversion cost | training gain |
+| `untuned_base` | the model before you trained it | *neither* — the difference confounds them | both |
+
+The headline table therefore comes from two runs, not one. `verify` reports the
+missing figure as `unavailable` with the reason, rather than deriving it from
+the two points it has — deriving it is the mistake the third point exists to
+prevent. Composing the three into a single command is the first thing on the
+list after this alpha; the README's *Limitations* records what is not wired.
+