relarena 0.0.1a1__tar.gz

relarena-0.0.1a1/LICENSE

@@ -0,0 +1,200 @@
+                                 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 PriorLabs GmbH
+
+   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

relarena-0.0.1a1/NOTICE

@@ -0,0 +1,68 @@
+relarena
+Copyright 2026 PriorLabs GmbH
+
+Licensed under the Apache License, Version 2.0. See the LICENSE file.
+
+This product includes third-party software. The following components are used
+under the MIT license, with the upstream copyright notice and license text
+retained alongside the code in this package:
+
+  * RelBench example GNN building blocks
+    Copyright (c) 2023 RelBench Team
+    https://github.com/snap-stanford/relbench
+    src/relarena/models/_shared/gnn/_vendor/gnn.py
+
+  * RelGNN composite-message-passing GNN
+    Copyright (c) 2023 RelBench Team
+    https://github.com/snap-stanford/RelGNN
+    src/relarena/models/relgnn/_vendor/
+
+  * RelGT Relational Graph Transformer
+    Copyright (c) 2025 Vijay Prakash Dwivedi, Sri Jaladi, Yangyi Shen,
+    Federico López, Charilaos I. Kanatsoulis, Rishi Puri, Matthias Fey,
+    Jure Leskovec
+    https://github.com/snap-stanford/relgt
+    src/relarena/models/relgt/_vendor/
+
+  * RDBLearn (method adaptation; no files copied)
+    Copyright (c) HKU Shanghai X-Lab
+    https://github.com/HKUSHXLab/rdblearn
+    src/relarena/models/rdblearn/model.py, src/relarena/models/_shared/tfm/tfm.py,
+    src/relarena/featurization/dfs.py
+
+  * Relational Transformer (RT) (dependency; no files copied)
+    Copyright (c) 2025 Stanford STAR / Relational Transformer authors
+    https://github.com/rishabh-ranjan/relational-transformer
+    src/relarena/models/rt/ -- installed from the `rt` extra, not vendored.
+
+Datasets are not distributed with this package. They are downloaded at runtime by
+relbench and remain subject to their own upstream terms.
+
+Vendored component provenance
+-----------------------------
+
+RelBench example GNN building blocks were copied from
+https://github.com/snap-stanford/relbench at commit 74d4c37
+(`examples/model.py` and `examples/text_embedder.py`). Their imports were merged
+into `src/relarena/models/_shared/gnn/_vendor/gnn.py`; only Ruff formatting was
+otherwise normalized. The GNN layers themselves are imported from the installed
+`relbench.modeling.nn` package and are not copied.
+
+RelGNN model building blocks were copied from
+https://github.com/snap-stanford/RelGNN at commit cffdb8b. The upstream example
+modules were split into the matching modules under
+`src/relarena/models/relgnn/_vendor/`; package imports were rewritten and Ruff
+formatting normalized. Duplicate RelBench encoders were not copied.
+
+RelGT model and sampler building blocks were copied from
+https://github.com/snap-stanford/relgt at commit 19e423ca. Upstream `utils.py`
+became `_sampler.py` and retains only the sampler functions and globals RelArena
+uses; inter-module imports were rewritten to package paths and Ruff formatting
+normalized. `models/relgt/model.py` and `models/relgt/tokenize.py` are original
+RelArena adapter code, not vendored files.
+
+RDBLearn is a method adaptation rather than a file copy. RelArena reimplements
+the Deep Feature Synthesis to tabular-foundation-model recipe from
+https://github.com/HKUSHXLab/rdblearn. The closest adaptation is the
+non-stratified train downsampling in `models/_shared/tfm/tfm.py`; RelArena's DFS
+cache and TabPFN-Rel extensions are original work.

relarena-0.0.1a1/PKG-INFO

@@ -0,0 +1,390 @@
+Metadata-Version: 2.4
+Name: relarena
+Version: 0.0.1a1
+Summary: A unified, fair benchmarking framework for models on relational tasks on RelBench databases.
+Keywords: benchmarking,machine-learning,relational-learning,relbench,tabular-data
+Author: Prior Labs
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+License-File: src/relarena/models/VENDORED-LICENSES
+Classifier: Development Status :: 3 - Alpha
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Dist: relbench==2.1.2
+Requires-Dist: numpy>=1.24
+Requires-Dist: pandas>=2.3.3,<3.0
+Requires-Dist: scikit-learn>=1.3
+Requires-Dist: configspace>=1.0
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: jsonschema>=4.0
+Requires-Dist: torch
+Requires-Dist: relarena[rdl] ; extra == 'graphsage'
+Requires-Dist: bencheval ; extra == 'leaderboard'
+Requires-Dist: lightgbm>=4.0,<4.7 ; extra == 'lightgbm'
+Requires-Dist: bencheval[plot] ; extra == 'plots'
+Requires-Dist: autorank ; extra == 'plots'
+Requires-Dist: tabpfn>=8 ; extra == 'rdblearn'
+Requires-Dist: fastdfs>=0.2.1 ; extra == 'rdblearn'
+Requires-Dist: setuptools<82 ; extra == 'rdblearn'
+Requires-Dist: torch-geometric>=2.5 ; extra == 'rdl'
+Requires-Dist: pytorch-frame>=0.2.3 ; extra == 'rdl'
+Requires-Dist: sentence-transformers ; extra == 'rdl'
+Requires-Dist: relarena[rdl] ; extra == 'relgnn'
+Requires-Dist: relarena[rdl] ; extra == 'relgt'
+Requires-Dist: einops>=0.8 ; extra == 'relgt'
+Requires-Dist: h5py>=3.0 ; extra == 'relgt'
+Requires-Dist: fastdfs>=0.2.1 ; extra == 'tabpfn-rel-api'
+Requires-Dist: setuptools<82 ; extra == 'tabpfn-rel-api'
+Requires-Dist: tabpfn-client>=0.3.2 ; extra == 'tabpfn-rel-api'
+Requires-Dist: relarena[rdblearn] ; extra == 'tabpfn-rel-local'
+Requires-Python: >=3.11, <3.13
+Project-URL: Homepage, https://github.com/PriorLabs/relarena
+Project-URL: Repository, https://github.com/PriorLabs/relarena
+Project-URL: Issues, https://github.com/PriorLabs/relarena/issues
+Provides-Extra: graphsage
+Provides-Extra: leaderboard
+Provides-Extra: lightgbm
+Provides-Extra: plots
+Provides-Extra: rdblearn
+Provides-Extra: rdl
+Provides-Extra: relgnn
+Provides-Extra: relgt
+Provides-Extra: tabpfn-rel-api
+Provides-Extra: tabpfn-rel-local
+Description-Content-Type: text/markdown
+
+# RelArena-α
+
+A unified, fair benchmarking framework for running models on relational tasks on
+[RelBench](https://github.com/snap-stanford/relbench) databases — inspired by how
+[TabArena](https://tabarena.ai) standardizes tabular benchmarking.
+
+This repository also open-sources TabPFN-Rel and an initial version of the
+Relational Predictive Interface (RPI). A detailed release report covering
+RelArena-α, TabPFN-Rel, and the RPI is in preparation.
+
+> **Current status:** alpha release. RelArena is a living benchmark: its task coverage,
+> baselines, API, and tuning regime will evolve with community feedback. The
+> current release focuses on RelBench v1 entity-level forecasting tasks.
+
+## Why
+
+Reproducibility varies across relational-learning methods: some releases omit
+training scripts or tuning details, and reported results often use different
+evaluation and tuning regimes. RelArena provides one executable
+*train → tune → evaluate* path with common data loading, split construction,
+model selection, and result recording. Models provide their training code and a
+declarative hyperparameter search space; callers choose the tuning budget.
+
+## Core idea
+
+```
+┌─ runner ───── fit config(s) on train → pick best on val → final fit → test
+├─ tuner ────── random search or a fixed grid under a caller-supplied budget;
+│               records configurations, metrics, predictions, and phase timings
+├─ model ────── RelArenaModel: fit / predict  (the contract)
+├─ space ────── SearchSpace: what to tune over, bound to the model in the registry
+└─ RelBench ─── Database, EntityTask, task.evaluate, metrics  (dependency)
+```
+
+**Evaluation protocol.** Each candidate configuration is fit on `train` and
+scored on `val`. The best validation configuration then receives a final fit and
+produces the test prediction. Depending on the model's published protocol, that
+final fit either refits on `train + val` or trains on `train` while retaining
+`val` for checkpoint selection. Parameter-free models simply run their sole
+configuration. Test labels are withheld from the model and supplied only to
+RelBench's evaluator.
+
+RelArena uses **nested temporal validation**: tuning receives a database censored
+at `val_timestamp`, while final evaluation receives one censored at
+`test_timestamp`. This prevents access to post-boundary data and test labels.
+Within that allowed database state, each method decides whether and how to enforce
+the finer timestamp of every historical example. See
+[docs/temporal-validation.md](docs/temporal-validation.md) for the complete
+guarantee and trade-off.
+
+Three design decisions carried over from TabArena/TabRepo:
+
+1. **The search space is decoupled from the model and declarative.** Models
+   implement just `fit` / `predict`; *what* to tune lives in a separate
+   `SearchSpace` (a `ConfigSpace.ConfigurationSpace` for random search, or an
+   explicit ordered `grid` for discrete spaces) bound to the model in the registry
+   — mirroring AutoGluon / TabArena rather than declaring the space on the class.
+2. **Budget is centralized rather than hidden in the model.** A model declares
+   only its `SearchSpace`; the caller supplies `n_trials`. The alpha release uses
+   documented, method-specific budgets because equalizing compute across methods
+   remains an open problem. See [docs/tuning-regime.md](docs/tuning-regime.md).
+3. **Runs retain useful metadata.** Each trial records its configuration, metrics,
+   optional predictions, and separate tuning/final-fit timings for later analysis.
+
+## Layout
+
+```
+src/relarena/
+  model.py        # RelArenaModel — the contract every model implements
+  search_space.py # SearchSpace — declarative HPO space (ConfigSpace or grid)
+  registry.py     # string-keyed model registry, binds model -> search space
+  tasks.py        # entity task-type scope + guard
+  metrics.py      # metric direction map + primary-metric selection
+  tuner.py        # random search / fixed grids; per-trial timing and predictions
+  runner.py       # local orchestration for one (model, dataset, task)
+  results.py      # TrialResult schema + DataFrame export
+  models/         # constant, lightgbm, rdblearn, graphsage, relgnn, relgt, tabpfn-rel, rt wrappers
+  featurization/  # relational DB -> flat feature table (entity-only, for now)
+  checksums/      # content fingerprints of the RelBench data + the recorded baseline
+  evaluation/     # leaderboard, plots, externally-reported reference baselines
+  userdb/         # Relational Predictive Interface (RPI)
+tests/            # smoke + unit tests (no data download)
+```
+
+## Models
+
+This is the canonical inventory of registered methods. The release snapshot and
+paper contain the rows marked **paper**; the additional `relgnn` registration
+is retained as an experimental final-fit variant.
+
+### Models and systems
+
+Every method registers as one of two kinds (`RelArenaModel.kind`), and the two
+are **not the same kind of result**. Both face the same tasks, splits, metrics,
+and runtime budget, so the comparison is fair on final performance: if a system
+scores higher, it really did do better on the benchmark. What a system gives up
+is the controlled setup. A **model** is one method under the harness's fixed
+tuning pipeline, so its score isolates the method. A **system** is free to step
+outside those constraints — it selects its own hyperparameters, training
+schedule, or components inside `fit` — so its score tells you what the whole
+package achieves without telling you which part earned it: how much comes from
+the underlying architecture rather than the selection machinery or other
+transferable tricks is not identifiable from the benchmark alone. Leaderboards
+should either exclude systems (`compute_leaderboard(..., kinds={"model"})`) or
+rank both populations together with systems clearly marked; publishing both
+boards side by side is the recommended presentation.
+
+System support is currently **highly experimental**: systems run through the ordinary
+model API with documented workarounds (all tuning inside a single fit of the
+default config, state carried between the fit and refit phases via module-level
+globals), and a system submission needs extra validation by and discussion with the maintainers. A
+future release will replace these workarounds with an explicit fitting API for
+systems — see
+[adding-a-model.md](docs/adding-a-model.md#model-or-system) for the current
+rules.
+
+| Registered identifier | Paper-facing name | Family | Kind | Status | Final fit | Extra |
+| --- | --- | --- | --- | --- | --- | --- |
+| `constant-global` | Constant (global) | global constant | model | paper | train + val | core |
+| `constant-per-entity` | Constant (per-entity) | entity-wise constant | model | paper | train + val | core |
+| `lightgbm` | LightGBM | entity-only tabular | model | paper | train + val | `lightgbm` |
+| `rdblearn` | RDBLearn | DFS + tabular foundation model | model | paper | train; val retained | `rdblearn` |
+| `tabpfn-rel-local` | TabPFN-Rel (OSS) | DFS + TabPFN v3 | model | paper | train + val | `tabpfn-rel-local` |
+| `tabpfn-rel-client` | TabPFN-Rel (API) | DFS + hosted TabPFN v3 with text | model | paper | train + val | `tabpfn-rel-api` |
+| `graphsage` | GraphSAGE | relational GNN | model | paper | train + val | `graphsage` |
+| `relgnn-es` | RelGNN | relational GNN | model | paper | best-validation checkpoint | `relgnn` |
+| `relgnn` | RelGNN full-data refit | relational GNN | model | experimental variant | train + val | `relgnn` |
+| `relgt` | RelGT | relational transformer | model | paper | best-validation checkpoint | `relgt` |
+| `rt-plurel` | RT-PluRel | pretrained relational transformer, fine-tuned per task | **system** | paper | train + val | `rt` |
+
+The paper reports `relgnn-es` simply as **RelGNN**, because that published-style
+best-validation-checkpoint regime performed better in our runs. The regular
+`relgnn` identifier remains available for experiments but is excluded from the
+default release leaderboard.
+
+RT-PluRel is the sole registered **system** (see
+[Models and systems](#models-and-systems)); its protocol and every configured
+value are documented in [`models/rt/model.py`](src/relarena/models/rt/model.py).
+Note that its recorded `val_score` is a placeholder — see
+[`baseline_results/README.md`](baseline_results/README.md).
+
+Everything else per method lives at its source: install caveats in
+[docs/adding-a-model.md §6](docs/adding-a-model.md#6-optional-dependencies)
+(the GNN baselines need platform-specific PyG sampling wheels beyond their
+extras), excluded backends and cache warmers in each model's docstring, and
+the complete implementation choices in the
+[adding-a-model appendix](docs/adding-a-model.md#appendix--what-every-existing-model-chose).
+
+## Install & test
+
+```bash
+uv sync            # the dev group (pytest, ruff, ...) installs by default
+OMP_NUM_THREADS=1 uv run pytest  # the prefix is required on macOS; harmless elsewhere
+```
+
+The `rt` integration is temporarily not exposed as a package extra because its
+platform-specific dependency is not yet on PyPI. To use it from a source
+checkout, uncomment the documented `rt` block in `pyproject.toml`, then run
+`uv lock` and `uv sync --extra rt`. Once `relational-transformer` is published
+on PyPI, the extra can use a normal version constraint without this local step.
+
+## Use RelArena on your own database (RPI)
+
+The **Relational Predictive Interface (RPI)** applies registered RelArena models
+to an entity-level forecasting task over your own relational database. Describe
+CSV or Parquet tables in a YAML database specification, define the forward-looking
+label and split boundaries in a YAML task specification, then use
+`PredictiveQuery` as the Python façade:
+
+```python
+from relarena.userdb import PredictiveQuery, PredictiveQuerySpec
+
+spec = PredictiveQuerySpec.from_yaml("task.yaml", data_dir="data/")
+predictions = PredictiveQuery(spec).fit("tabpfn-rel-client").predict()
+```
+
+See [docs/predictive-task.md](docs/predictive-task.md) for the task definition,
+SQL rules, split semantics, and worked examples.
+
+## Preprocessing caches (optional)
+
+Expensive CPU-bound preprocessing may run before a benchmark and be reused across
+trials. This can substantially reduce repeated-run time, especially for DFS feature
+matrices, materialized graphs, and tokenized databases.
+
+Caching is not required. RelArena provides an **optional, experimental** helper API
+in [`relarena.cache`](src/relarena/cache.py) for local paths, miss policies, private
+scratch computation, and atomic publication. A method may ignore this API and
+implement caching independently. The helper does not bring cache warming into a
+timed RelArena experiment; preprocessing scripts still run separately, so their
+runtime is not currently included in the recorded experiment timings.
+
+Regardless of the mechanism, cache-generation code must be public, reproducible,
+and leakage-safe. Some practical pointers:
+
+- Load data through `RelBenchDatasetTask.inner_split()` and `outer_split()` so the
+  validation and test phase boundaries remain intact.
+- Let the preprocessing implementation own its keys, versions, serialization, and
+  validation; include only inputs that actually determine the artifact.
+- Treat pre-built stores as a convenience: always ship a runnable warmer that can
+  reconstruct them.
+
+The full implementation guidance and reference code live in
+[docs/adding-a-model.md](docs/adding-a-model.md#4-pre-processing-cache).
+
+Configure the store explicitly at the run entrypoint:
+
+```python
+run_experiment(..., cache_dir="~/relarena-cache")
+```
+
+Entrypoints also resolve these environment variables once:
+
+- `RELARENA_CACHE_DIR` — the store directory.
+- `RELARENA_DISABLE_CACHE` — set to any value to disable persistent caches.
+
+`RELARENA_DISABLE_FEATURE_CACHE` remains as a deprecated alias for one release.
+The helper API's store is an ordinary local directory; remote snapshot transport
+belongs to deployment infrastructure rather than RelArena itself.
+
+**Precompute (CPU).** Because `fit` / `predict` read (a miss raises), build the store
+up front with a fill run. The DFS engine (`fastdfs`) runs on CPU and is memory-hungry
+on wide-fan-out schemas, so run it on a large CPU node (many cores, ample RAM);
+everything after it runs on the GPU (or the hosted TabPFN API), so precomputing keeps
+that CPU-heavy step off those nodes. The workflow warms every RelBench v1 task:
+
+```bash
+RELARENA_CACHE_DIR=~/relarena-cache \
+    uv run --extra rdblearn python workflows/warm_feature_cache.py
+```
+
+It invokes `relarena.featurization.warm_cache` for both protocol splits and warms
+both legitimate outer histories: train-only for RDBLearn and train+val for models
+that refit on all labeled data. The `tabpfn-rel` and `rdblearn`
+models share full-anchor, leak-safe-history matrices whenever their actual inputs
+match; model-specific row selection and downstream training do not affect the key.
+On a warm cache the evaluation reads Parquet only — no RDB build and no DFS.
+RelGNN, RelGT, and RT-PluRel expose independent runnable warmers at
+`relarena.models.relgnn.warm_cache`, `relarena.models.relgt.warm_cache`, and
+`relarena.models.rt.warm_cache`.
+
+**Runnable demo.** `examples/tabpfn_rel_caching.py` fits one RelBench task with and
+without a precomputed cache, reports both timings, and checks the outputs are
+identical. Its header includes a CPU-only mode (`RELARENA_EXAMPLE_SKIP_TFM=1`)
+that exercises the DFS and cache path without a GPU.
+
+## Batch evaluation
+
+The CLI runs one model across many tasks in-process and writes every evaluated
+config to a CSV:
+
+```bash
+relarena --model lightgbm --datasets rel-f1 --output results.csv
+```
+
+Each `(model, dataset, task, seed)` experiment is independent, so sweeps
+parallelize trivially. The building blocks are all public — `run_experiment`
+executes one experiment, `summary_to_dataframe` flattens it into the shared
+results schema, and concatenated frames feed the leaderboard (needs the
+`leaderboard` extra):
+
+```python
+import pandas as pd
+
+import relarena.models  # registers the built-in models
+from relarena.evaluation import compute_leaderboard
+from relarena.registry import registry
+from relarena.results import summary_to_dataframe
+from relarena.runner import run_experiment
+from relarena.tasks import list_entity_tasks
+
+frames = []
+for spec in list_entity_tasks(["rel-f1"]):
+    for model in ("constant-global", "lightgbm"):
+        summary = run_experiment(
+            registry.get(model), spec.dataset, spec.task, seed=0, n_trials=10
+        )
+        frames.append(summary_to_dataframe(summary))
+board = compute_leaderboard(pd.concat(frames, ignore_index=True))
+```
+
+If you have a large-scale cluster, integrate that loop into your distributed
+backend of choice (a SLURM array, Ray, ...): dispatch each experiment as one
+job, cache each job's result frame keyed by `(model, dataset, task, seed,
+n_trials)`, and concatenate the cached frames for the leaderboard. Warm the
+shared caches first (`workflows/warm_feature_cache.py` and the per-model
+`warm_cache` modules) so workers never pay the featurization cost.
+
+## Baseline results
+
+[`baseline_results/`](baseline_results/) holds the release snapshot of the sweep
+over the RelBench-v1 entity tasks: `results.csv` (every evaluated config; feed
+it to `compute_leaderboard`) and `reference_results.csv` — per-task scores for
+methods **not reproduced in this pipeline**, transcribed from published model
+reports and flagged with a `_MR` (**model report**) suffix. `_MR` numbers are
+mostly self-reported and are often higher than the results reproduced through
+RelArena; we discuss possible reasons in the forthcoming release report
+of this README. With few exceptions, these results are
+not directly comparable to RelArena runs and should only be used as reference
+points. To include them in a leaderboard or plot, pass `reference=`
+(`relarena.evaluation.load_reference_results`). See
+[`baseline_results/README.md`](baseline_results/README.md) for per-method
+provenance and caveats.
+
+## Adding a model
+
+A model is a folder under `src/relarena/models/` implementing the
+`RelArenaModel` contract (`fit` / `predict`) with a `SearchSpace` registered via
+`@register_model(search_space=...)`; the registry discovers the folder
+automatically. `models/lightgbm/` is the smallest complete example to copy.
+
+The full guide is [docs/adding-a-model.md](docs/adding-a-model.md): the layout,
+the datatypes `fit` and `predict` receive, the tuning regime and its choices,
+where shared code goes, optional dependencies, vendoring requirements, tests,
+and a checklist.
+
+## License
+
+Apache-2.0 ([`LICENSE`](LICENSE), [`NOTICE`](NOTICE)). Two things the license on
+this code does not settle, both worth reading before you rely on relarena:
+
+- **`tabpfn` is not Apache-2.0.** It ships the Prior Labs License, an Apache-2.0
+  derivative whose added paragraph 10 requires anyone distributing a product built
+  on it to display "Built with PriorLabs-TabPFN". It is confined to the `rdblearn`
+  and `tabpfn-rel-*` extras, so a plain install does not pull it.
+- **Datasets are not ours to license.** relarena serves no data itself; `relbench`
+  downloads every database at runtime, and they remain subject to their own
+  upstream terms.
+
+See [`docs/licensing.md`](docs/licensing.md) for what the license does and does not
+cover, and [`NOTICE`](NOTICE) for third-party attribution.