sdm-learn 0.1.0.dev0__tar.gz → 0.1.2__tar.gz

sdm_learn-0.1.2/CHANGELOG.md

@@ -0,0 +1,96 @@
+# Changelog
+
+All notable changes to `sdm-learn`. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions follow
+[PEP 440](https://peps.python.org/pep-0440/).
+
+## [0.1.2] - 2026-09-09
+
+### Added
+- `sdm ui` is now the whole product: a **Runs** table over everything under
+  `runs/` (filters, checkboxes), **New run**, **Run detail** with the document
+  epoch by epoch, and **Compare**, which opens on every finished run grouped
+  by dataset and switches formats and gate rules like the showcase page.
+  Deep links: `#run/<name>`, `#compare/<a>,<b>`, `#new`.
+- `sdm showcase OUT.html [GROUP/]LABEL[|RULE]=RUN_DIR ...` writes the
+  standalone comparison page from the same component (`sdm.showcase`);
+  `reproduce/showcase.py` is gone.
+
+## [0.1.1] - 2026-09-09
+
+### Added
+- `sdm ui`: a local web page that configures and launches runs (dataset,
+  algorithm, format, style, provider / model / API key, gate settings) and
+  shows the live log, the per-epoch verdicts, the results, and the learned
+  document. Standard library only; binds to localhost.
+- `reproduce/showcase.py` takes several datasets in one page
+  (`GROUP/LABEL=RUN_DIR`) with a dataset switcher.
+- `rulebook-html` format: the numbered rulebook written out as an HTML page
+  with a section per class, and a `SKETCH | class | <svg>` edit that gives a
+  class one schematic. `--passes N` walks the training rows N times;
+  `--max-rules 100,200` sets one cap per pass; `--no-gate` keeps every valid
+  revision. The four science datasets default to this protocol, which is
+  the one their checked-in rulebooks were learned under.
+
+## [0.1.0] - 2026-09-09
+
+The first non-development release. The library was cut down to what the
+paper describes and made to run through one command.
+
+### Added
+- `sdm run --dataset NAME|DIR --algo sdm|base --format ...`: one command for
+  every benchmark and for a directory of your own `train.jsonl` /
+  `test.jsonl` rows (`src/sdm/benchmarks.py`). Each built-in dataset carries
+  the gate settings its protocol uses; flags override them.
+- `--style` / `train(style=...)`: a free-text guide, shown only to the
+  writer, naming the kinds of content the document should use.
+- SEARCH/REPLACE `<patch>` revisions for the `text`, `html`, `image`, and
+  `audio` formats, so a revision costs its difference rather than a rewrite.
+- A `text` format: one plain-text document.
+- `image` format refuses every text element; the picture carries the
+  hypothesis.
+- Gateway knobs: `SDM_TEMPERATURE`, `SDM_EXTRA_PAYLOAD`,
+  `SDM_MAX_COMPLETION_TOKENS`, and a separate writer model or endpoint
+  (`SDM_WRITER_MODEL`, `SDM_WRITER_EFFORT`, `SDM_WRITER_GATEWAY_URL`,
+  `SDM_WRITER_API_KEY`).
+- Writer robustness for small models: lenient edit parsing, one corrected
+  retry on an invalid revision, context-window failures skip the candidate,
+  `proposal_examples` caps the feedback shown per epoch.
+- `reproduce/showcase.py`: an interactive page comparing formats on one
+  dataset (accuracy or PR-AUC, document growth, per-class curves, every
+  epoch's document with playable audio).
+- Four science benchmarks with their learned HTML rulebooks and designed
+  views: `gravityspy`, `galaxy10`, `bloodmnist`, `eclipsing` (`sdm.science`,
+  the `science` extra, `reproduce/ai4science`).
+- The offline test suite is tracked and run by CI.
+
+### Changed
+- **Algorithms are `sdm` and `base`** (formerly `gate` and `baseline`).
+- **The gate accepts ties.** A candidate is kept unless it lowers protected
+  accuracy; the strict-gain rule and the 5% compression exception are gone.
+  On PODS this raised the markdown rulebook from 91.7 to 95.4 PR-AUC.
+- **`cycles` are `epochs`** everywhere: options, state, candidate log, raw
+  folders. Runs saved under the old name do not resume.
+- `format.md` is the single source of the default formats
+  (`DEFAULT_FORMAT_MD` reads it). Size caps: `html` 12,000 words / 60,000
+  characters, `image` 40,000 characters.
+- Rendered formats ask the writer for ASCII text (cairosvg has no glyph for
+  arrows and draws a box).
+- README rewritten as the package's reference documentation.
+
+### Removed
+- The in-context (`icl`) control, the GEPA and SkillOpt baselines, the
+  fine-tuning export, the graded free-response mode, and the ungated
+  `accept_all` switch.
+- Every dataset loader and experiment except `cifar10`, `pods`, `df2`, and
+  `camelyon17` (GSM8K, MBPP, LiveMath, IFBench, tabular, Omniglot, Dogs, the
+  other WILDS subsets, and more).
+- The `video` format: on PODS, 70 objects in 12 frames reached 18.5 PR-AUC
+  against 64.5 for one wordless picture.
+- The `--val` protected-validation option of the Camelyon17 command
+  (`holdout_split` remains in the Python API).
+
+## [0.1.0.dev0] - 2026-08-29
+
+First upload: the gate, the `markdown`, `html`, `image`, `audio`, and
+`video` formats, CIFAR-10, and the experiment template.

sdm_learn-0.1.2/FAQ.md

@@ -0,0 +1,24 @@
+# Frequently Asked Questions
+
+## Where do I configure the LLM?
+
+Copy `settings.example.json` to `settings.json`, then set `provider`, `model`,
+and `api_key` there. `settings.json` is gitignored.
+
+## Which LLM APIs are supported?
+
+See the **VIEW SUPPORTED APIs** table in the
+[README](README.md#configure-an-llm-api). OpenAI-compatible providers can also
+be configured with `"provider": "custom"` and a `gateway_url`.
+
+## Where should I put my experiment?
+
+Copy `runs/my_experiment.py`, update its `load_data()` function, and run the
+copy from the repository root. The shipped benchmarks need no copy:
+`sdm run --dataset cifar10|pods|df2|camelyon17`. Files created under `runs/` are gitignored except for
+the template.
+
+## Why use a virtual environment?
+
+It keeps this project's dependencies isolated and ensures the package is
+installed with the Python interpreter you intend to use.

sdm_learn-0.1.2/MANIFEST.in

@@ -0,0 +1,12 @@
+include LICENSE
+include README.md
+include FAQ.md
+include CHANGELOG.md
+include settings.example.json
+
+# The GIFs stay out of the distribution: 2.5 MB of animation helps a reader on
+# GitHub and does nothing for an install. The README reaches them by relative
+# path, which GitHub resolves and PyPI does not, so the PyPI page shows a
+# broken image until that link is made absolute at release time.
+prune assets
+prune runs