LightPFN 1.0.0__tar.gz

lightpfn-1.0.0/.gitignore

@@ -0,0 +1,16 @@
+__pycache__/
+*.pyc
+*.egg-info/
+dist/
+build/
+.pytest_cache/
+.venv/
+# training and evaluation outputs
+data/
+runs/
+*.pt
+*.safetensors
+# LaTeX build files (the PDF is tracked)
+paper/*.aux
+paper/*.log
+paper/*.out

lightpfn-1.0.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 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 [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
+
+       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.

lightpfn-1.0.0/NOTICE

@@ -0,0 +1,11 @@
+LightPFN
+Copyright 2026 Giorgio Ottoboni
+
+This product includes software and model weights licensed under the Apache License, Version 2.0.
+
+The model was trained from scratch on synthetic data only, without distillation and without weights or
+outputs of other tabular foundation models.
+
+The optional training code (lightpfn.prior) uses the structural causal model prior of TabICL
+(https://github.com/soda-inria/tabicl, BSD-3-Clause, Copyright (c) 2025, Soda team @ Inria) as an installed
+dependency; TabICL code is not redistributed in this repository or in the published wheel.

lightpfn-1.0.0/PKG-INFO

@@ -0,0 +1,298 @@
+Metadata-Version: 2.5
+Name: LightPFN
+Version: 1.0.0
+Summary: A 4.6M-parameter tabular classifier pretrained only on synthetic data, with a scikit-learn API and CPU, CUDA/ROCm and Vulkan inference
+Project-URL: Homepage, https://github.com/GioOtto/LightPFN
+Project-URL: Documentation, https://github.com/GioOtto/LightPFN/blob/main/docs/en/GUIDE.md
+Project-URL: Paper, https://github.com/GioOtto/LightPFN/blob/main/paper/LightPFN_report.pdf
+Project-URL: Weights, https://huggingface.co/ueuegio/LightPFN
+Project-URL: Issues, https://github.com/GioOtto/LightPFN/issues
+Author-email: Giorgio Ottoboni <247403232+GioOtto@users.noreply.github.com>
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Keywords: classification,foundation model,in-context learning,scikit-learn,tabpfn,tabular,vulkan
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: huggingface-hub>=0.27
+Requires-Dist: numpy>=1.24
+Requires-Dist: safetensors>=0.5
+Requires-Dist: scikit-learn>=1.6
+Requires-Dist: torch>=2.6
+Provides-Extra: dev
+Requires-Dist: build>=1.2; extra == 'dev'
+Requires-Dist: pandas>=2.0; extra == 'dev'
+Requires-Dist: pytest>=8; extra == 'dev'
+Provides-Extra: eval
+Requires-Dist: catboost>=1.2; extra == 'eval'
+Requires-Dist: joblib>=1.3; extra == 'eval'
+Requires-Dist: lightgbm>=4.0; extra == 'eval'
+Requires-Dist: openml>=0.15; extra == 'eval'
+Requires-Dist: pandas>=2.0; extra == 'eval'
+Requires-Dist: threadpoolctl>=3.1; extra == 'eval'
+Requires-Dist: xgboost>=2.0; extra == 'eval'
+Provides-Extra: train
+Requires-Dist: pandas>=2.0; extra == 'train'
+Requires-Dist: scipy>=1.11; extra == 'train'
+Requires-Dist: tabicl==2.2.0; extra == 'train'
+Requires-Dist: threadpoolctl>=3.1; extra == 'train'
+Provides-Extra: vulkan
+Requires-Dist: wgpu<0.33,>=0.32; extra == 'vulkan'
+Description-Content-Type: text/markdown
+
+<div align="center">
+  <img src="https://raw.githubusercontent.com/GioOtto/LightPFN/main/docs/assets/logo.svg" width="112" height="112" alt="LightPFN logo: a small table with one highlighted cell" />
+  <h1>LightPFN</h1>
+  <p><strong>A 4.6M-parameter tabular foundation model for classification.</strong></p>
+  <p>Pretrained only on synthetic data. A scikit-learn classifier that runs on CPU, CUDA, ROCm and any Vulkan GPU.</p>
+
+  <p>
+    <a href="https://pypi.org/project/LightPFN/"><img alt="PyPI" src="https://img.shields.io/pypi/v/lightpfn?style=for-the-badge&color=111111&label=PyPI" /></a>
+    <a href="https://huggingface.co/ueuegio/LightPFN"><img alt="Weights on Hugging Face" src="https://img.shields.io/badge/Weights-Hugging%20Face-111111?style=for-the-badge&logo=huggingface&logoColor=white" /></a>
+    <a href="https://github.com/GioOtto/LightPFN/blob/main/paper/LightPFN_report.pdf"><img alt="Technical report (PDF)" src="https://img.shields.io/badge/Technical%20report-PDF-8B1A1A?style=for-the-badge&logo=adobeacrobatreader&logoColor=white" /></a>
+  </p>
+
+  <p>
+    <a href="https://github.com/GioOtto/LightPFN/blob/main/docs/en/GUIDE.md">User guide</a> &nbsp;&nbsp;
+    <a href="https://github.com/GioOtto/LightPFN/blob/main/docs/en/RESULTS.md">Results</a> &nbsp;&nbsp;
+    <a href="https://github.com/GioOtto/LightPFN/blob/main/docs/en/VULKAN.md">Vulkan backend</a> &nbsp;&nbsp;
+    <a href="https://github.com/GioOtto/LightPFN/blob/main/docs/en/TRAINING.md">Training</a> &nbsp;&nbsp;
+    <a href="https://github.com/GioOtto/LightPFN/issues">Issues</a>
+  </p>
+  <p><img alt="Python 3.10+" src="https://img.shields.io/badge/python-3.10%2B-181818" /> <img alt="4.6M parameters" src="https://img.shields.io/badge/parameters-4.6M-181818" /> <img alt="Code and weights Apache 2.0" src="https://img.shields.io/badge/code%20%2B%20weights-Apache%202.0-181818" /> <img alt="Trained only on synthetic data" src="https://img.shields.io/badge/training%20data-synthetic%20only-181818" /></p>
+  <p><strong>English</strong> &nbsp; <a href="https://github.com/GioOtto/LightPFN/blob/main/README.it.md">Italiano</a></p>
+</div>
+
+## What it is
+
+LightPFN is a prior-data fitted network: a transformer pretrained once on millions of synthetic
+classification tasks. At `fit` it stores your training set as context; at `predict_proba` it reads that
+context and the test rows in one forward pass. It does not train on your data and requires no tuning.
+
+```python
+from lightpfn import LightPFNClassifier
+
+clf = LightPFNClassifier(n_estimators=4).fit(X_train, y_train)
+proba = clf.predict_proba(X_test)
+```
+
+- **Small.** 4,603,088 parameters, 18 MB of weights, designed for a commodity CPU. TabICLv2 has six
+  times as many.
+- **Accurate without tuning.** On 55 OpenML datasets outside TabArena, mean AUC 0.911: 0.86 points above
+  default CatBoost and 1.8 to 2.1 points above default LightGBM, XGBoost and random forest. On the 38
+  TabArena classification tasks, lower error than default CatBoost on 76% of the tasks.
+- **Any GPU.** CUDA and ROCm through PyTorch, and a Vulkan backend with its own WGSL compute kernels for
+  AMD, Intel and NVIDIA GPUs on Linux and Windows, with no CUDA or ROCm install. On an RX 7900 XT it is
+  6 to 10 times faster than a 16-thread CPU.
+- **scikit-learn API.** `Pipeline`, `GridSearchCV`, `clone`, `cross_val_score`, and pandas DataFrames
+  with categorical, string, boolean and missing values.
+- **Open.** Code, weights and the full training pipeline under Apache 2.0. Trained from scratch: no
+  distillation, no weights or outputs of other tabular foundation models.
+
+## Benchmarks
+
+**TabArena-Lite, official pipeline, 38 classification datasets, 99 methods** (8 October 2026).
+LightPFN uses its default configuration with four estimators, eight-fold bagging and the official validation
+protocol. All 38 tasks succeeded, none imputed. **25th of 99, Elo 1420 (+67 / -66)**.
+
+| Method | Elo | 95% CI |
+|---|---:|---:|
+| TabICLv2 (default) | 1558 | +77 / -68 |
+| TabDPT-1.3 (default) | 1467 | +78 / -54 |
+| RealMLP (tuned + ensembled) | 1459 | +50 / -47 |
+| **LightPFN (default)** | **1420** | **+67 / -66** |
+| CatBoost (tuned) | 1378 | +58 / -55 |
+| CatBoost (tuned + ensembled) | 1370 | +58 / -48 |
+| LightGBM (tuned + ensembled) | 1365 | +52 / -42 |
+| XGBoost (tuned + ensembled) | 1346 | +58 / -62 |
+| CatBoost (default) | 1339 | +49 / -52 |
+| XGBoost (default) | 1191 | +57 / -69 |
+| LightGBM (default) | 1144 | +59 / -63 |
+| RandomForest (default) | 1000 | +71 / -84 |
+
+<p align="center"><img src="https://raw.githubusercontent.com/GioOtto/LightPFN/main/docs/assets/tabarena_lite_pareto.png" width="100%" alt="TabArena-Lite: Elo against fit time, LightPFN highlighted among foundation models" /></p>
+
+The point estimate is above every GBDT in this leaderboard, including tuned and ensembled entries; the intervals
+with tuned CatBoost overlap, so this does not establish a clear win. TabICLv2 and larger foundation models
+lead. Median fit time is 0.65 s per 1,000 rows on the RX 7900 XT; among entries with at least its Elo, only
+TabDPT-1.3 has a lower reported time. Leaderboard timings come from different hardware. This is an author-run
+Lite evaluation; TabArena maintainers re-run the full benchmark before leaderboard inclusion.
+[Protocol, results and artifacts](https://github.com/GioOtto/LightPFN/blob/main/docs/en/RESULTS.md).
+
+## Install
+
+```bash
+pip install LightPFN             # CPU, CUDA or ROCm, through your PyTorch install
+pip install "LightPFN[vulkan]"   # adds the Vulkan GPU backend (wgpu)
+```
+
+Python 3.10 or newer. The first `fit` downloads the weights (18 MB) from
+[Hugging Face](https://huggingface.co/ueuegio/LightPFN) at a pinned commit and caches them. On a CPU-only
+machine, install PyTorch from its CPU index first to skip the CUDA libraries:
+`pip install torch --index-url https://download.pytorch.org/whl/cpu`.
+
+## Quick start
+
+```python
+from sklearn.datasets import load_breast_cancer
+from sklearn.metrics import roc_auc_score
+from sklearn.model_selection import train_test_split
+
+from lightpfn import LightPFNClassifier
+
+X, y = load_breast_cancer(return_X_y=True)
+X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y, random_state=0)
+
+clf = LightPFNClassifier(n_estimators=4, random_state=0)
+clf.fit(X_train, y_train)
+print(roc_auc_score(y_test, clf.predict_proba(X_test)[:, 1]))
+```
+
+A pandas DataFrame with categorical and missing values goes in as it is:
+
+```python
+import pandas as pd
+
+df = pd.DataFrame({
+    "age": [34, 51, None, 28, 45, 39],
+    "plan": ["basic", "pro", "pro", None, "basic", "enterprise"],
+    "region": pd.Categorical(["north", "south", "south", "east", "north", "east"]),
+    "active": [True, False, True, True, False, True],
+})
+y = [0, 1, 1, 0, 1, 0]
+clf = LightPFNClassifier().fit(df, y)
+clf.predict_proba(df.head(2))
+```
+
+Columns of category, string, object or bool dtype become ordinal codes of the categories seen in `fit`.
+A category dtype keeps its declared levels, including unused ones. Missing values and values outside
+that vocabulary become NaN, which the model handles natively. More in
+[examples/](https://github.com/GioOtto/LightPFN/tree/main/examples/) and in the [user guide](https://github.com/GioOtto/LightPFN/blob/main/docs/en/GUIDE.md).
+
+## Devices
+
+| `device=` | Runs on |
+|---|---|
+| `"auto"` (default) | a CUDA or ROCm GPU through PyTorch, else a Vulkan GPU, else the CPU |
+| `"cpu"` | the CPU, through PyTorch |
+| `"cuda"`, `"cuda:1"` | an NVIDIA (CUDA) or AMD (ROCm) GPU through PyTorch |
+| `"vulkan"`, `"vulkan:1"` | any GPU with a Vulkan driver, through `lightpfn.vulkan` (needs `lightpfn[vulkan]`) |
+
+The environment variable `LIGHTPFN_DEVICE` replaces `"auto"`. Every device returns the same probabilities
+up to floating-point rounding (within 3e-5). See the [Vulkan backend](https://github.com/GioOtto/LightPFN/blob/main/docs/en/VULKAN.md).
+
+## When to use it
+
+LightPFN is a good default for classification tables with 2 to 10 classes and up to tens of thousands of
+rows, when you want a strong model without tuning. Know its limits:
+
+- **Classification only**, 2 to 10 classes. Regression comes with version 2.
+- **Large tables.** Above `max_context` rows (default 20,000) each estimator reads a stratified subsample.
+  In the report, with the whole training set as context, its mean AUC is level with default CatBoost
+  up to 100,000 rows. Larger contexts cost more; sizes beyond 100,000 were not evaluated.
+- **Categorical columns** are read as ordinal codes. On tables dominated by high-cardinality categorical
+  columns, CatBoost is ahead (by up to 4.5 AUC points on two TabArena tasks).
+- **CPU time grows with the context.** On TabArena, four estimators take a median of 1 s per split on small
+  tables, 15 s on medium and 54 s on large ones, against 3, 4 and 7 s for CatBoost. When time matters, use a
+  GPU or `n_estimators=1` (about four times faster, slightly less accurate).
+
+## Results
+
+Baselines at default settings. Differences in AUC points (100 times the AUC difference) with 95% paired
+bootstrap intervals. Full tables, per-task results and timings: [docs/en/RESULTS.md](https://github.com/GioOtto/LightPFN/blob/main/docs/en/RESULTS.md).
+
+**55 OpenML-CC18 datasets outside TabArena** (at most 1,000 rows and 100 features, five-fold
+cross-validation, one estimator):
+
+| Model | Mean AUC | LightPFN minus model |
+|---|---:|---:|
+| **LightPFN** | **0.911** | |
+| CatBoost | 0.902 | +0.86 [0.42, 1.40] |
+| Random forest | 0.893 | +1.76 |
+| LightGBM | 0.891 | +2.02 |
+| XGBoost (54 datasets: its wrapper fails on one) | 0.889 | +2.06 [1.32, 2.94] |
+
+**TabArena, 38 classification tasks** (official splits, first repeat, run with our own harness, which is
+not the official leaderboard protocol):
+
+| Model | Mean AUC | Mean rank of 7 | Lower error than CatBoost |
+|---|---:|---:|---:|
+| TabICLv2 (28M parameters, GPU) | 0.864 | 1.58 | 87% |
+| **LightPFN, 4 estimators** | **0.858** | **2.50** | **76%** |
+| LightPFN, 1 estimator | 0.857 | 3.37 | 68% |
+| CatBoost | 0.855 | 3.50 | |
+| LightGBM | 0.844 | 5.34 | 5% |
+| Random forest | 0.838 | 5.89 | 3% |
+| XGBoost | 0.833 | 5.82 | 11% |
+
+## How it works
+
+<p align="center"><img src="https://raw.githubusercontent.com/GioOtto/LightPFN/main/docs/assets/architecture.png" width="100%" alt="LightPFN architecture: cell embedding, two column stages, row refinement and compression, in-context learning, retrieval decoder" /></p>
+
+Each cell is embedded from its value, its rank in the column and a missing-value flag. Two column stages
+with induced attention read the labelled context of each column. A row stage with four summary tokens mixes
+the features of each row and compresses it into one vector. Seven in-context blocks let every test row
+attend to the training rows, and a retrieval decoder turns the attention into a vote over the training
+labels.
+
+Pretraining used 7.68 million task draws from 4.03 million distinct synthetic tasks: 90% from a structural causal graph prior and 10% from a rule
+prior (XOR, parity, lookup tables, trees) that teaches feature interactions. A second stage trained on
+tables of up to 60,000 rows. The [technical report](https://github.com/GioOtto/LightPFN/blob/main/paper/LightPFN_report.pdf) describes the model, the
+priors and the selection protocol; [docs/en/TRAINING.md](https://github.com/GioOtto/LightPFN/blob/main/docs/en/TRAINING.md) explains how to reproduce
+the training.
+
+## Roadmap
+
+Version 2 will add:
+
+- **Native categorical features.** The model will see the categorical mask from the start of pretraining,
+  and the prior will contain high-cardinality categorical columns whose many rare levels each carry a small
+  effect. An adapter trained on top of the frozen v1 weights did not help, because the v1 prior has no such
+  columns to learn from (report, Section 12).
+- **Regression.**
+
+## Repository layout
+
+| Path | Content |
+|---|---|
+| `lightpfn/` | the package: model, scikit-learn wrapper, devices, Vulkan backend (`vulkan/`), and the training code: priors (`prior/`), trainer (`train.py`), evaluation harness (`eval/`) |
+| `tests/` | unit and equivalence tests (CPU; the Vulkan tests also run on a CPU driver) |
+| `examples/` | runnable examples |
+| `docs/` | user guide, results, Vulkan backend and training in English (`en/`) and Italian (`it/`), changelog, third-party licenses |
+| `paper/` | technical report: PDF, LaTeX source, figures and plot data |
+
+The published wheel contains only the inference code; training needs a source checkout
+(`pip install -e ".[train,eval]"`).
+
+## Contributing and support
+
+Bug reports and focused pull requests are welcome: see [CONTRIBUTING](https://github.com/GioOtto/LightPFN/blob/main/.github/CONTRIBUTING.md). Security
+issues go through [SECURITY](https://github.com/GioOtto/LightPFN/blob/main/.github/SECURITY.md). Changes between versions are in
+[CHANGELOG](https://github.com/GioOtto/LightPFN/blob/main/docs/CHANGELOG.md).
+
+## Citation
+
+```bibtex
+@techreport{ottoboni2026lightpfn,
+  title  = {A Sling Against Giants: {LightPFN}, a 4.6M-parameter tabular in-context classifier designed to stay small},
+  author = {Ottoboni, Giorgio},
+  year   = {2026},
+  url    = {https://github.com/GioOtto/LightPFN}
+}
+```
+
+GitHub's "Cite this repository" button reads [CITATION.cff](https://github.com/GioOtto/LightPFN/blob/main/CITATION.cff).
+
+## License
+
+Code and weights: [Apache License 2.0](https://github.com/GioOtto/LightPFN/blob/main/LICENSE), with the attribution notice in [NOTICE](https://github.com/GioOtto/LightPFN/blob/main/NOTICE).
+Dependencies keep their own licenses, listed in [THIRD_PARTY_LICENSES](https://github.com/GioOtto/LightPFN/blob/main/docs/THIRD_PARTY_LICENSES.md).